Soil health assessment method based on multi-source data fusion

Abstract

The invention relates to the technical field of soil health assessment, and discloses a soil health assessment method based on multi-source data fusion. The method comprises the following steps: collecting physical, chemical and biological index data of soil in real time through a sensor and monitoring equipment, constructing a health index matrix and normalizing the health index matrix to obtain initial health matrix data; an evaluation request is received, a health description vector is generated through analysis, a health urgency factor is calculated, and health demand data is obtained; traversing the initial health matrix data according to the health description vector by using an improved clustering algorithm to perform coarse-grained matching to obtain candidate health state subset data; secondly, optimizing the subset data through an adaptive neural network model, selecting a state of maximizing health rewards in combination with a reward mechanism, and outputting an optimal health state to generate an evaluation instruction; and finally, evaluating soil health according to an instruction, monitoring optimal health state real-time data, and starting a health data calibration protocol when the optimal health state real-time data exceeds a preset threshold value, thereby realizing effective fusion and dynamic accurate evaluation of multi-source data.

Description

Technical Field

[0001] This invention relates to the field of soil health assessment technology, specifically a soil health assessment method based on multi-source data fusion. Background Technology

[0002] Soil, as a crucial carrier of agricultural production and ecosystem stability, has its health status directly linked to food security and environmental quality. With the increasing demands of modern agriculture and ecological protection, the need for soil health assessment is becoming increasingly urgent; however, many problems remain to be addressed in the current soil health assessment process.

[0003] Traditional soil health assessments often rely on single types of data, such as collecting only chemical indicators like soil pH and organic matter content, or focusing solely on physical indicators like soil bulk density and porosity. They lack effective integration of soil biological indicators, such as microbial community structure and enzyme activity. This single-data-source approach fails to comprehensively reflect the true state of soil health, easily leading to biased assessment results and failing to provide a comprehensive basis for soil management decisions. Existing assessment methods have significant shortcomings in data processing and matching. Most methods use traditional clustering algorithms for data matching. These algorithms are poorly adaptable to data distribution and struggle to achieve accurate coarse-grained matching when faced with noise, redundancy, and magnitude differences between different indicators in multi-source data. They often filter out a large number of irrelevant or redundant candidate health status data, increasing the difficulty and computational cost of subsequent fine-grained optimization. Existing technologies also fall short in terms of the dynamic adjustment and calibration of assessment results. Soil health status changes in real time with factors such as changes in the external environment and agricultural production activities, but most current assessment methods are static assessments, lacking an effective monitoring mechanism for real-time data on optimal health status. Even if thresholds are set, it is difficult to quickly initiate a scientific and reasonable health data calibration protocol when the data exceeds the threshold, resulting in assessment results that cannot be updated in a timely manner and cannot adapt to the dynamic changes in soil health status, thus affecting the timeliness and effectiveness of soil management measures. Summary of the Invention

[0004] The purpose of this invention is to provide a soil health assessment method based on multi-source data fusion to solve the problems mentioned in the background art.

[0005] To achieve the above objectives, the present invention provides a soil health assessment method based on multi-source data fusion, the method comprising: Multi-source soil data is collected, and soil physical, chemical and biological indicators are acquired in real time through sensors and monitoring equipment. A health indicator matrix is ​​constructed based on the multi-source soil data and normalized to obtain initial health matrix data. Receive a soil health assessment request, parse the soil health assessment request to generate a health description vector, and calculate the health urgency factor to obtain health demand data; Based on the improved clustering algorithm, coarse-grained matching is performed by traversing the initial health matrix data according to the health description vector to obtain a subset of candidate health status data. The candidate health state subset data is fine-grained by using an adaptive neural network model, and the state that maximizes the health reward is selected through a reward mechanism to output the optimal health state and obtain the health assessment instruction. Soil health is assessed according to the health assessment instructions, and real-time data of the optimal health status is monitored. When the real-time data of the optimal health status exceeds a preset threshold, a health data calibration protocol is initiated.

[0006] Preferably, the collection of multi-source soil data includes: Soil physical, chemical, and biological indicators were collected to obtain multi-source soil data. The physical indicators included soil moisture, soil temperature, and soil density. The chemical indicators included pH value, organic matter content, and nutrient element concentration. The biological indicators included microbial activity and enzyme activity. Based on the multidimensional indicators of each monitoring point in the multi-source soil data, a health indicator matrix is ​​established in the order of the monitoring points. The health indicators are normalized using a standard normalization method, so that the indicators of each dimension are within the same range of values, thus obtaining the initial health matrix data.

[0007] Preferably, receiving a soil health assessment request includes: Receive a soil health assessment request, obtain the assessment parameters, health standards and time requirements in the soil health assessment request, and obtain the health assessment parameters; The health assessment parameters are integrated into a health description vector. The current time and the time requirement in the health description vector are obtained. The health urgency factor is calculated based on the pre-set weight coefficients and health level.

[0008] Preferably, the coarse-grained matching based on the improved clustering algorithm includes: Based on the assessment parameters and health standards in the health description vector, the initial health matrix data is traversed to select monitoring points that meet the basic requirements of health assessment, and the monitoring points are combined into a first candidate health status subset. The first candidate health status subset is sorted from high to low according to the health index scores of the monitoring points to obtain the second candidate health status subset; A health status group is established based on an improved clustering algorithm, with the left side being the health description vector and the right side being the second candidate health status subset; An adjustment variable is introduced to address multiple health constraints, and an optimization function is established. By using group center and boundary search, the top candidate health status is output, and a subset of candidate health status data is obtained.

[0009] Preferably, the fine-grained optimization of the candidate health state subset data using an adaptive neural network model includes: Define the state space, action space, and reward function of the adaptive neural network model; The real-time state of the candidate health state subset data is used as the input as the state space of the model. The action space is used to generate binary actions of selection and non-selection for each state of the candidate health state subset, and outputs a single optimal health state. The reward function is determined to include a multi-dimensional dynamic reward mechanism, which includes at least positive and negative rewards.

[0010] Preferably, the state that maximizes health rewards through the reward mechanism includes: Based on the subset of candidate health status data, status data of candidate health status is collected, and actions are selected through an exploration strategy. The exploration strategy includes at least two approaches: probabilistically exploring insufficiently evaluated states and probabilistically selecting the state with the highest current reward value. Calculate the immediate reward based on the actual results, update the reward value using a value update formula, output the optimal health status, and obtain the health assessment instruction.

[0011] Preferably, the real-time data for monitoring the optimal health status includes: Combine the optimal health status identifier with health assessment parameters into an assessment instruction; Use a health assessor to verify the current availability of health data; If the verification passes, the health assessment process will be initiated; if the verification fails, the data anomaly handling process will be triggered. The system has multiple preset threshold levels, including warning thresholds and emergency thresholds. The warning threshold is defined as soil moisture deviation exceeding a first preset deviation tolerance and pH value deviation exceeding a second preset deviation tolerance for a duration of a first preset duration. The emergency threshold is defined as soil moisture deviation exceeding a third preset deviation tolerance and pH value deviation exceeding a fourth preset deviation tolerance for a duration of a second preset duration. The third preset deviation tolerance is greater than the first preset deviation tolerance, the fourth preset deviation tolerance is greater than the second preset deviation tolerance, and the second preset duration is shorter than the first preset duration. When the real-time data of the optimal health status exceeds the preset threshold of any level, the health data calibration protocol is activated.

[0012] Preferably, the initiation of the health data calibration protocol includes: The calibration amount is calculated based on the deviation between real-time health data and standard health data. Based on the calibration values, adjust the health assessment parameters and regenerate the health description vector; The optimal health status is updated by performing coarse-grained matching and fine-grained optimization using the regenerated health description vector.

[0013] Preferably, the collected soil physical, chemical, and biological data include: Using a soil sensor network to continuously monitor physical parameters, including soil moisture, soil temperature, and soil density; Use a chemical analyzer to regularly measure chemical indicators, including pH value, organic matter content, and nutrient concentration; Use biosensors to detect biological indicators, including microbial activity and enzyme activity; Integrate all data to generate multi-source soil data.

[0014] Preferably, the calculation of the health urgency factor includes: Based on the time requirements and health standards in the health assessment parameters, and combined with the current time, the time urgency is calculated; Health importance is calculated based on health level and weighting coefficient. By combining time urgency with health importance, a health urgency factor is generated.

[0015] Compared with the prior art, the beneficial effects of the present invention are: By integrating data from three key soil indicators—physical, chemical, and biological—this method overcomes the limitations of traditional assessment methods that rely on a single data source. Traditional methods often focus on only one or two types of indicators, failing to comprehensively reflect the complex state of soil health. In contrast, this method collects multi-source data and constructs a health indicator matrix. Normalization then eliminates interference caused by differences in magnitude between different indicators, enabling the initial health matrix data to more comprehensively and objectively reflect the fundamental information of soil health. This provides richer and more reliable data support for subsequent assessments, avoiding assessment biases caused by incomplete data. In the data matching stage, an improved clustering algorithm is used for coarse-grained matching. Compared with the traditional clustering algorithm, the improved algorithm is more adaptable to multi-source data and can better handle noise and redundancy in the data. When traversing the initial health matrix data, it can more accurately filter out the subset of candidate health states related to the health description vector, effectively reducing the mixing of irrelevant data, reducing the computational burden in the subsequent fine-grained optimization process, improving the efficiency of the overall evaluation process, and laying a good foundation for subsequent accurate optimization. The application of adaptive neural network models provides an efficient solution for fine-grained optimization of candidate health state subset data. This model possesses the ability to learn and adjust autonomously, continuously optimizing the analysis process based on the characteristics of health demand data and candidate data. Combined with a reward mechanism, it selects the state that maximizes health rewards, accurately identifying the optimal health state and significantly improving the accuracy of health assessment results. This makes the assessment results more closely reflect the actual situation of soil health, providing a more precise reference direction for soil management decisions. This method establishes a mechanism for real-time monitoring of optimal soil health status and the activation of a health data calibration protocol, enabling real-time tracking of changes in soil health status. When the real-time data of optimal soil health status exceeds a preset threshold, the calibration protocol can be activated promptly to adjust the assessment data and results. This breaks through the limitations of traditional static assessments, allowing soil health assessments to dynamically adapt to changes in soil state caused by external environmental factors and agricultural activities. It ensures that assessment results maintain high timeliness and effectiveness, facilitating timely adjustments to soil management measures based on changes in soil health status, and better meeting the needs of agricultural production and ecological protection for soil health management. Attached Figure Description

[0016] Figure 1 This is a schematic diagram illustrating the working principle of the soil health assessment method based on multi-source data fusion described in this invention. Figure 2 A flowchart for multi-source soil data acquisition and processing; Figure 3 A flowchart for improving coarse-grained matching in clustering algorithms. Detailed Implementation

[0017] 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, and 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.

[0018] Please see Figure 1This invention provides a soil health assessment method based on multi-source data fusion. The method includes: collecting multi-source soil data; acquiring soil physical, chemical, and biological indicators in real time through sensors and monitoring equipment; constructing a health indicator matrix based on the multi-source soil data and performing normalization processing to obtain initial health matrix data; receiving a soil health assessment request; parsing the request to generate a health description vector and calculating a health urgency factor to obtain health demand data; performing coarse-grained matching based on an improved clustering algorithm by traversing the initial health matrix data according to the health description vector to obtain a subset of candidate health states; performing fine-grained optimization on the subset of candidate health states using an adaptive neural network model; selecting the state that maximizes the health reward through a reward mechanism; outputting the optimal health state to obtain a health assessment instruction; assessing soil health according to the health assessment instruction; monitoring the real-time data of the optimal health state; and initiating a health data calibration protocol when the real-time data of the optimal health state exceeds a pre-set threshold.

[0019] Example 1: See Figure 2 Multi-source soil data acquisition was accomplished through a distributed sensor network. Physical index monitoring employed buried soil moisture sensors, temperature probes, and density meters. These devices automatically collected topsoil data at preset frequencies. The moisture sensors, based on the dielectric constant principle, provided real-time feedback on volumetric water content. The temperature probes recorded diurnal variation curves at a depth of 20 cm below the surface. Density measurement utilized gamma-ray attenuation to obtain profile data. Chemical index acquisition relied on a mobile chemical analysis unit equipped with a pH electrode, an organic matter spectral analysis module, and an ion-selective electrode array. The sampling cycle was dynamically adjusted according to the crop growth stage. pH measurement was performed in situ in the field using a potentiometry method. Organic matter content was calculated using near-infrared spectral reflectance, and nutrient element concentrations were determined through electrochemical analysis of soil extracts. Biological index capture utilized a biosensor chip. Microbial activity detection was based on carbon dioxide flux generated by respiration, while enzyme activity was measured using a fluorescence substrate method to measure hydrolytic enzyme reaction rates. All biosensors were equipped with a temperature control module to maintain optimal reaction conditions.

[0020] The integration of multi-source soil data occurs on a central data processing platform. This platform receives raw data streams from different monitoring nodes, first performing spatiotemporal alignment by mapping discrete monitoring points to a unified grid coordinate system using GPS coordinates, and then synchronously generating minute-level data slices using timestamps. The health indicator matrix is ​​constructed according to a two-dimensional structure of "monitoring point × indicator dimension". The matrix row vectors correspond to fixed geographic coordinates, and the column vectors are arranged in the order of physical, chemical, and biological indicators. The physical indicator sub-column contains three sets of time-series data: humidity, temperature, and density. The chemical indicator sub-column stores parameters such as pH value, organic matter, and nitrogen, phosphorus, and potassium concentrations. The biological indicator sub-column records microbial respiration intensity and activity values ​​such as phosphatase and urease. Normalization is performed using a deviation standardization method, calculating the range for each column of indicator data separately, and linearly transforming the original values ​​to the [0,1] interval. Humidity data is normalized to two decimal places, and low-value indicators such as enzyme activity are converted using scientific notation to ensure that the dimensions of each dimension are consistent and comparable. The parsing module for soil health assessment requests is deployed on a cloud server. When it receives an assessment instruction in JSON format, the parsing engine automatically extracts three core fields: assessment parameters specifying the target crop type and growth stage, health standards specifying the threshold ranges for each indicator, and time requirements limiting the deadline for outputting the assessment results. The health assessment parameters are encapsulated using vectorization, generating a 32-dimensional floating-point health description vector. The first 8 dimensions encode the weights of physical indicators, the middle 12 dimensions correspond to the priorities of chemical indicators, and the last 12 dimensions carry the constraints of biological indicators. Vector element values ​​are automatically filled according to the predefined scoring rules in the standard document. The calculation of the health urgency factor activates the time management module. This module compares the system's current clock with the time requirement field in the request. When the remaining time is less than 20% of the total time, a time urgency coefficient (value 1.5) is triggered. The health importance coefficient is determined by querying a preset weight table; for example, the weight for the critical growth period of grain crops is set to 2.0. The final factor value is the product of the time coefficient and the health coefficient.

[0021] The generation process of the health description vector includes a dynamic verification mechanism. An integrity check is performed immediately after vector initialization, missing parameters are automatically filled with default values, and conflicting parameters trigger a negotiation protocol. For example, when organic matter content requirements conflict with nutrient element standards, the system prioritizes the main crop's requirements. The sorting logic of vector elements follows an indicator priority rule: soil moisture always takes precedence among physical indicators, pH value takes precedence over nutrients among chemical indicators, and microbial activity is ranked higher than specific enzyme activity among biological indicators. This sorting method ensures that core parameters receive higher computational weight. The application of the health urgency factor is reflected in the assessment process scheduling level. High-urgency requests are automatically redirected to the fast assessment channel, which uses a simplified clustering algorithm and shortens the neural network training cycle, while triggering a resource preemption mechanism to suspend low-priority tasks.

[0022] The sensor network deployment scheme considers the spatial heterogeneity of soil. In typical farmland, an equidistant grid layout is used, with 9 monitoring nodes per hectare. In hilly areas, sampling density is increased according to terrain units. Physical sensors employ industrial-grade protective designs. Humidity probes are equipped with self-cleaning electrodes to avoid the influence of salt crystallization, temperature sensors are equipped with radiation shields to eliminate sunlight interference, and density probes undergo depth calibration to compensate for formation pressure errors. The chemical analysis unit implements a strict quality control process, performing three-point calibration before each batch of sampling. The pH electrode is calibrated daily with standard buffer solution, and the spectral analysis module updates the local crop model database weekly. The innovative design of the biosensor chip enables in-situ detection. The microbial activity detection chamber integrates a gas exchange membrane, and the enzyme activity reaction chamber uses microfluidic technology to control the substrate injection volume. All biological data acquisition processes are maintained at a constant temperature of 37℃±0.5℃. The calculation of time urgency incorporates nonlinear processing. When the ratio of remaining time to total time is less than 0.3, an acceleration mode is activated, and the urgency coefficient increases exponentially, reaching a maximum value of 3.0 when the ratio is below 0.1. The importance of health is quantified using a tiered weighting method. Food security crops are assigned a basic weight of 1.8, cash crops 1.5, and green manure crops 1.2. The weights for crop growth stages are configured as follows: seedling stage 0.8, jointing stage 1.2, heading stage 1.5, and maturity stage 1.0. The health urgency factor is synthesized using matrix multiplication. The time vector and the health weight vector are multiplied by a dot product to generate a scalar factor, which serves as a meta-parameter controlling the iteration depth and computational intensity of subsequent evaluation processes.

[0023] The data processing platform implements a multi-level caching mechanism. Raw sensor data first enters the edge computing nodes for preliminary filtering to remove pulse interference signals caused by equipment failure. After receiving the preprocessed data, the central server performs secondary cleaning, using a sliding window method to identify outliers. The cleaning window for soil moisture data is set to 5 minutes, while enzyme activity data, due to its slow changes, uses a 30-minute window. The health indicator matrix employs a block storage strategy for memory management. Physical indicator blocks are allocated contiguous memory space to support fast traversal, chemical indicator blocks use sparse matrix compression storage, and biological indicator blocks have a time-series index for convenient historical data backtracking. The parallel computing architecture for normalization processing significantly improves efficiency. The physical indicator group is allocated an independent computing unit to perform real-time normalization, while the chemical and biological indicator groups share the processing core but adopt a time-sharing reuse strategy. The entire normalization pipeline ensures that data updates from 200 monitoring points are processed per minute. The semantic analysis module in the parameter parsing process supports natural language input. When receiving unstructured request text, the semantic engine automatically extracts key phrases and maps them to a standard parameter set. For example, "key water requirement stage during corn tasseling" is parsed as crop type = corn, growth period = tasseling stage, and core indicator = soil moisture. The dynamic update function of the health description vector allows for mid-process correction. When a sudden rainfall event is detected, the system automatically increases the weight of the humidity indicator and generates a revised version of the vector. The feedback adjustment mechanism of the health urgency factor achieves dynamic balance. When the assessment process times out, the factor value is automatically increased to trigger an acceleration mode. If the assessment is completed ahead of schedule, the factor value is reduced to release computing resources. This flexible design effectively copes with the unpredictability of the farmland environment. The entire implementation process establishes a closed-loop quality control system. The sensor network is equipped with a self-diagnostic function to report the equipment status in real time. The chemical analysis unit has a built-in standard sample verification module. The biosensor sets up a positive control group. The data processing platform implements version control to record each algorithm update. The assessment results must pass a consistency check before output, ensuring traceability of the entire chain from data collection to health assessment. This implementation method demonstrated stable performance in tests in typical black soil areas, successfully handling complex evaluation requests with minute-level data update frequencies, and providing technical support for precision agricultural management.

[0024] Example 2: See Figure 3The initial stage of coarse-grained matching initiates a data traversal engine. This engine loads the evaluation parameters and health standard fields from the health description vector. The evaluation parameters include specific constraints such as a target pH range of 6.0-7.2 and an organic matter content threshold of 3.5%, while the health standards define the allowable fluctuation range for each indicator. The traversal operation scans each monitoring point record in the initial health matrix, extracting data from each row and comparing it with the health standards. When the pH value of a monitoring point exceeds the 7.2 upper limit three times consecutively, a removal flag is triggered, and records with organic matter content below 3.0% are automatically moved to the waiting queue for review. The screening process generates a first candidate health status subset, which retains the numbers of all monitoring points that meet the basic constraints, while also recording the number of violations for each point as a basis for subsequent ranking.

[0025] The sorting module for the first candidate subset employs a multi-key value fast sorting algorithm, with the primary sorting key being the health indicator score. This score is obtained through weighted calculation: physical indicators have a weight of 0.3, chemical indicators 0.5, and biological indicators 0.2. Soil moisture compliance rate accounts for 60% of the physical score, and pH stability accounts for 40% of the chemical score. The sorting process establishes a min-heap data structure, extracting the top 50 monitoring points with the highest scores each time to form a temporary ordered list. When two monitoring points have the same score, their variance over the most recent 24 hours is compared, with the smaller variance receiving higher priority. The final generated second candidate health status subset contains 200 optimal monitoring points, stored in a circular buffer in descending order of score. The buffer has an automatic overwrite mechanism to retain the latest sorting results. An improved clustering algorithm initializes the grouping structure. The left group center is dynamically generated from the health description vector, mapping the 32-dimensional parameters in the vector to 32-dimensional spatial coordinates. The right group loads the monitoring point data from the second candidate subset, with each monitoring point corresponding to a location in space. The grouping process incorporates three adjustment variables: α controls pH matching accuracy (value 0.05-0.2), β regulates organic matter content tolerance (value 1%-3%), and γ manages biomarker weights (value 0.1-0.5). The optimization function is defined as a composite objective of minimizing Euclidean distance and maximizing intra-group similarity, with a distance weight of 0.7 and a similarity weight of 0.3. The function is solved using simulated annealing, with the temperature parameter decreasing by 10% every 100 iterations, starting from 1000.

[0026] The group center update process performs iterative calculations. The first round of calculation uses the health description vector coordinates as the initial center, scanning the coordinates of all monitoring points on the right. Points within a radius R are grouped into a temporary group, with R initially set to 0.5. The mean values ​​of the monitoring points within each group are calculated to generate a new center point. Regrouping is triggered when the distance between the new and old centers exceeds a threshold δ. The δ value is dynamically adjusted based on the health urgency factor: δ = 0.1 for high urgency and δ = 0.05 for normal conditions. Boundary search uses a convex hull algorithm to determine the grouping range, constructing a minimum boundary polyhedron in 32-dimensional space. Monitoring points on the boundary undergo secondary screening, eliminating outliers deviating from the center by more than two standard deviations. The output mechanism for the top candidate health status is designed as a sliding window mode, extracting the top K optimal monitoring points from the optimized group each time. The K value is determined by the health urgency factor: K = 10 when the factor > 2.0, K = 20 when the factor is between 1.0 and 2.0, and K = 30 when the factor < 1.0. The extraction process records the matching degree between the spatial coordinates of each monitoring point and the health description vector. The matching degree is calculated using the cosine similarity method, with higher weighting coefficients assigned to the physical indicator dimension. The candidate health status subset data is stored in a four-tuple structure: monitoring point ID, health score, group number, and matching degree index. Data packets are transmitted to the fine-grained optimization module via a message queue, and data verification is implemented during transmission to prevent packet loss.

[0027] The algorithm is deployed on a distributed computing cluster, with each computing node responsible for processing group calculations for 50 monitoring points. Nodes synchronize center point coordinates via the MPI protocol. Memory management employs a paging storage strategy, with health description vectors residing in a high-speed memory cache, and monitoring point data stored in different memory pages according to group numbers. An iterative calculation process includes a timeout interrupt mechanism; if a single iteration takes more than 200 milliseconds, it automatically jumps to the next round to prevent individual anomalies from causing overall computational stagnation. The grouping result visualization module generates a 3D projection map, reducing the 32-dimensional space to the main indicator plane, allowing operators to monitor the group distribution status in real time. An adaptive mechanism for adjusting variables connects to external environmental sensors. When the rainfall sensor detects precipitation >10mm, the weight coefficient of the soil moisture index is automatically increased; when the temperature sensor records sustained high temperatures >35℃, the tolerance for microbial activity indicators is increased by 20%. The dynamic parameter configuration interface for the optimization function allows agronomists to manually intervene, adjusting the weight ratio of distance and similarity via sliders; modifications take effect immediately without restarting the calculation process. The boundary search algorithm integrates an anomaly detection module, which initiates special diagnosis for monitoring points that are at the edge of the group three times in a row, and generates a health assessment report containing 12 detection indicators.

[0028] The persistent storage of candidate health status subset data employs a time-series database. Each grouping result is indexed by timestamp, retaining complete calculation records for the most recent 30 days. A dedicated binary format is defined for data transmission, with packet headers containing data version identifiers and checksums. The payload utilizes zlib compression to reduce bandwidth consumption. A double-buffered receiving area is set up at the receiving end. When the current grouped data enters the processing buffer, newly arriving data is temporarily stored in the reserve buffer to ensure computational continuity. The historical comparison function for health status grouping supports backtracking analysis, allowing overlay display of group distribution changes across different time periods to aid in judging soil health evolution trends. A quality monitoring system is established throughout the coarse-grained matching process. Data integrity checks are performed before grouping calculations, and the standard deviation of monitoring points within a group is calculated after grouping. Regrouping is triggered when the standard deviation of physical indicators exceeds 0.15. Computational resource allocation is dynamically scheduled; high-priority evaluation requests can occupy additional computing nodes, while regular tasks use a time-slice round-robin mechanism. In a typical application scenario, this implementation successfully processed a large farmland dataset containing 5000 monitoring points, with an average matching time controlled within 3 seconds.

[0029] Example 3: The adaptive neural network model is constructed using a hierarchical architecture. The state space is defined as a multi-dimensional feature representation of the candidate health state subset data. Each state includes a monitoring point number, a timestamp, and standardized indicator values. The physical indicator dimension retains minute-level sampling values ​​of soil moisture, temperature, and density; the chemical indicator dimension stores hourly average values ​​of pH, organic matter content, and nitrogen, phosphorus, and potassium concentrations; and the biological indicator dimension records the daily fluctuation range of microbial activity and enzyme activity. The action space is designed using a discrete binary selection mechanism. For each candidate state, a decision instruction to select or abandon is generated. The selection action triggers the state to enter the actuarial process, while the abandonment action marks the state as evaluated and releases computing resources. The reward function architecture integrates multi-source feedback signals. Positive rewards are derived from the proximity of health indicators to target values, while negative rewards are related to resource consumption and time costs. The reward value calculation incorporates a health urgency factor as a dynamic adjustment coefficient.

[0030] The model training process is deployed on a GPU-accelerated cluster. During initialization, historical data from a subset of candidate healthy states are loaded. A 32-dimensional feature vector is extracted from each state as input. The first hidden layer has 128 neurons using the ReLU activation function, and the second hidden layer has 64 neurons using the Sigmoid activation function. The output layer generates the probability distribution of two actions. An ε-greedy algorithm is implemented for the exploration strategy. Initially, the exploration rate is set to 0.7 to encourage extensive attempts at unknown states. After every 1000 decisions, the exploration rate is decayed by 5%, remaining constant when it falls below 0.1. The action selection mechanism establishes a value evaluation system. Each state maintains a Q-value representing the expected long-term cumulative reward. When selecting an action, the action with the largest Q-value is chosen with a probability of (1-ε), and states that have not been sufficiently evaluated are randomly selected with a probability of ε.

[0031] The calculation of immediate rewards employs a multi-dimensional dynamic evaluation. After an action is selected, the system monitors the changes in indicators over the next 3 hours. For every 1% deviation of soil moisture from the target value, 0.2 reward points are deducted; maintaining pH within the optimal range for 1 hour increases the reward by 0.5 points; and a 10% increase in microbial activity awards 1.0 point. Reward value updates utilize a temporal difference learning algorithm. The new Q-value is calculated as a weighted combination of the current reward and the maximum Q-value for the next state.

[0032] in: Indicates the current state. Indicates the execution of an action. For instant rewards, This is the subsequent state. For subsequent actions, The learning rate is set to 0.05. The discount factor is set to 0.9. An early stopping mechanism is set for model training, which automatically terminates the iteration when the reward fluctuation is less than 0.01 for 20 consecutive training rounds.

[0033] The optimal health state output process includes a verification step. After the neural network generates an initial selection, it compares the matching degree between this state and the health description vector. When the cosine similarity is below 0.6, a reselection is triggered. Health assessment instructions are generated using a standardized message format. The instruction header includes a timestamp, assessment version number, and urgency level identifier. The instruction body specifies the monitoring point number of the optimal state, the values ​​of various indicators, and the expected duration of maintenance. Additional fields record key decision parameters from the selection process. Instruction transmission uses a reliable transmission protocol. The assessment execution process is initiated after the receiving end returns a confirmation signal. The online learning function of the neural network model supports real-time optimization. After each health assessment, the actual effect is compared with the predicted reward. When the reward error exceeds a threshold, model parameter fine-tuning is triggered. The fine-tuning process uses mini-batch gradient descent, with a batch size of 32 historical state samples and a learning rate reduced to one-tenth of the initial value to prevent overfitting. Model version management is strictly controlled. Before each update, the effect is verified in a test environment. New models must achieve an accuracy of over 90% on the validation set before being deployed to the production system.

[0034] State feature engineering employs deep processing. Raw index data undergoes moving average filtering to eliminate random noise; soil moisture data uses first-order difference to extract trends; pH sequences are identified through Fourier transform to recognize periodic patterns; and bioactivity indicators are normalized to eliminate dimensional influences. A feature selection algorithm calculates the correlation coefficients between each dimension and the health score, retaining 16 core features with correlation coefficients greater than 0.3. Redundant features are automatically removed to reduce computational complexity. Feature vectors are ultimately scaled to a uniform numerical range: physical indicators are mapped to the [0,1] interval; chemical indicators are standardized using z-scores; and biological indicators undergo logarithmic transformation to improve distribution symmetry. The exploration strategy optimization incorporates a Boltzmann distribution, calculating action selection probabilities based on Q-values. The temperature parameter τ is initially set to 1.0 and gradually decreased during training, until it drops to 0.1, at which point it relies entirely on the action with the highest Q-value. This design promotes exploration diversity in the early stages of training and enhances the utilization of known optimal strategies in later stages. A state access counter records the number of times each candidate state is selected. When a state is accessed more than 50 times, its exploration priority is automatically reduced to avoid excessive resource concentration on a few states. The dynamic adjustment mechanism of the reward function is connected to the external environment perception system. When the weather forecast indicates rainfall within the next 6 hours, the reward weight of the soil moisture index is increased; when crops are detected to have entered a critical growth stage, the reward score of the nutrient element index is increased. The reward calculation also considers the time factor; behaviors that quickly reach a healthy state receive additional time rewards, while decisions that take longer than the average time are penalized accordingly. Reward value normalization maps all rewards to the [-1,1] interval to ensure training stability.

[0035] The model deployment employs a distributed inference architecture, with multiple candidate states computing Q-values ​​in parallel via neural networks. GPU tensor cores accelerate matrix operations, keeping the processing time for a single state within 5 milliseconds. A decision result caching mechanism stores the most recent 1000 selection records, prioritizing cached results to reduce computation when encountering similar states. A system monitoring panel displays key metrics such as exploration rate, average reward, and decision accuracy in real time. Model retraining is automatically triggered when the average reward of 10 consecutive decisions falls below a threshold. A feedback loop is established throughout the fine-grained optimization process. After each health assessment, actual performance data is collected, the difference between predicted and actual rewards is calculated, and this data is fed back to the training system. This difference data is used to calibrate the reward function parameters, and reward function reconstruction is initiated when systematic deviations occur over multiple cycles. In six months of actual operation, this implementation successfully processed over 12,000 health assessment requests, improving model decision accuracy from an initial 68% to a stable 92%, and reducing the average decision time to 1.2 seconds.

[0036] Example 4: The monitoring system continuously tracks the monitoring point with the optimal health status identifier "CornField-789", located in a corn-growing area at 38°N latitude. The assessment instruction requires maintaining soil moisture within the healthy range of 22%±2% and pH value of 6.8±0.3. The real-time data stream is updated every 5 minutes. The data acquisition module recorded that at 14:00 on July 12, soil moisture suddenly dropped to 18.6% and pH value fluctuated to 7.2. The system immediately activated the threshold comparison engine. The warning threshold setting parameters show that a soil moisture deviation tolerance of ±3% for 120 minutes triggers a warning, and a pH deviation tolerance of ±0.5 for 90 minutes triggers a warning; the emergency threshold requires a moisture deviation exceeding ±5% or a pH deviation exceeding ±0.8 for 60 minutes to trigger an alarm. Currently, the moisture deviation has reached 3.4% for 15 minutes and the pH deviation has reached 0.4 for 20 minutes. The system marks the status as "early warning observation" but does not immediately trigger calibration.

[0037] When the health assessor performs data availability verification, it first checks the sensor status codes. The humidity probe returns a calibration flag "0x1A" (indicating calibration completed within 3 days), and the pH electrode status code is "0x2F" (indicating no calibration within 7 days). The verification logic requires that the calibration cycle of key indicator sensors not exceed 5 days; therefore, the pH electrode triggers a "data reliability question" flag. The system automatically switches to the backup electrode data source, which records a pH value of 7.1 (deviation 0.3). After verification, the assessment execution process is initiated. If the backup source is unavailable, an exception handling is triggered: the current assessment instruction is frozen, and a mobile testing vehicle is dispatched to the site for retesting. At 15:30 on July 12, data showed that the humidity remained below 19.2% for 90 minutes, and the pH value rose to 7.4 for 45 minutes. The system detected that two warning thresholds were simultaneously exceeded: humidity deviation 3.8% > 3% tolerance and continued to exceed the limit, and pH deviation 0.6 > 0.5 tolerance. The threshold cross-validation module confirmed that the "warning threshold condition combination" was met, and the health data calibration protocol was immediately initiated. The calibration calculation engine uses 7 days of historical health data to generate standard reference curves: humidity baseline 22.3% ± 0.5 (daily fluctuation range), pH baseline 6.82 ± 0.15. The deviation between real-time data and baseline values ​​is calculated by integration: humidity calibration ΔH = +3.7% (increased irrigation required), pH calibration ΔP = -0.58 (acidic amendment required).

[0038] Table 1: Multi-level Threshold Trigger Parameter Configuration Table Threshold level Soil moisture deviation tolerance pH deviation tolerance Duration requirement Related response mechanism Warning threshold >±3% >±0.5 Lasting 90 minutes Activate secondary monitoring frequency Emergency threshold >±5% >±0.8 Lasting 60 minutes Trigger Instant Calibration Protocol Disaster threshold >±8% >±1.2 Lasting 30 minutes Activate the emergency response system After receiving calibration data, the health assessment parameter adjustment module modifies the original assessment parameters: the target humidity range is adjusted from 22%±2% to 22%±1.5% (narrowing the tolerance band), and the pH target is adjusted from 6.8±0.3 to 6.8±0.2. The parameter revision log records the modification time, operator (automatically by the system), and revision reason code "THRESHOLD_BREACH". When the new health description vector is regenerated, the weight of physical indicators is increased from 0.35 to 0.42, the weight of chemical indicators is decreased from 0.45 to 0.40, and the weight of biological indicators remains unchanged at 0.18. The vector version number is upgraded from V1.2 to V1.3, and the time requirement field is compressed to 70% of the original duration. After restarting coarse-grained matching, an optimized path is adopted, the clustering algorithm center point is shifted to the low humidity area by 0.7 standard deviations, and the candidate subset selection condition adds "humidity change rate in the last 3 hours" as a new dimension. In the fine-grained optimization phase, the neural network reward function increases the humidity stability reward weight by 30%, and the pH reward function introduces a secondary penalty term (the penalty coefficient doubles when the deviation exceeds 0.5). A new optimal state identifier, "CornField-789B," is output after re-evaluation. This state requires precise drip irrigation at 2.5 cubic meters per hour, along with the application of 150 kg of sulfur powder per hectare. The calibration protocol execution process involves equipment linkage. The system opens the solenoid valve in area 789 via the IoT gateway. The irrigation volume controller receives the calibration amount of ΔH = +3.7%, converting it into a replenishment volume of 37.5 cubic meters (based on the soil volumetric moisture content conversion model). The pH calibration command triggers the fertilizer applicator premixing system, calculating the citric acid solution injection volume based on ΔP = -0.58, and adding a pH adjuster to a concentration of 5.8 in the irrigation water. The execution progress is fed back to the monitoring center every 5 minutes. When the humidity rises to 20.1% at 15:45, the system lowers the alarm level but continues to operate the calibration protocol. Closed-loop validation of data calibration was initiated 2 hours after irrigation. The system collected stratified profile samples: 0-20cm layer with 21.7% humidity and pH 6.9, and 20-40cm layer with 22.1% humidity and pH 6.7. Validation rules required that the deviation of each layer's indicators from the target value should not exceed 15% of the calibration amount, and the measured data met the validation standards (humidity deviation 0.6% < 0.55%, pH deviation 0.1 < 0.087). Upon successful calibration, the health status database was updated, the health score of the monitoring point recovered from 82 to 92, the system deactivated the alarm, and a calibration report was generated and archived.

[0039] The entire monitoring process employs a three-tiered redundancy mechanism. The main monitoring system utilizes real-time stream processing, the backup system performs snapshot comparisons every 10 minutes, and the disaster recovery system retains a complete hourly data mirror. Threshold parameters are dynamically managed: the lower humidity threshold is automatically relaxed by 0.5% during the rainy season, and the pH fluctuation tolerance is tightened by 0.1% during the hot season. Calibration protocol version control records the details of each adjustment, and automatically initiates a comprehensive sensor verification procedure when calibration is triggered more than 5 times in a single month. In the 2024 dry season test in northern China, the system successfully handled 127 threshold alarm events, with an average calibration command response time controlled within 18 minutes, effectively maintaining the stability of soil health.

[0040] Example 5: The soil sensor network was deployed using a grid-based strategy. Thirty-six monitoring nodes were set up in a 15-hectare experimental field in the Northeast Black Soil region, with a hexagonal distribution and a 25-meter spacing between nodes. Each node had three layers of sensors: a capacitive humidity probe and a PT100 temperature sensor installed 10 cm below the surface; a gamma-ray densitometer at a depth of 20 cm; and an auxiliary humidity calibration point at a depth of 40 cm. Physical index data collection followed an adaptive frequency adjustment principle: data was collected every 30 minutes in sunny weather, automatically increasing to every 5 minutes during rainfall. The minimum resolution for temperature data recording was 0.1℃. Density measurement used a dual-probe differential method to eliminate the influence of compaction. Data transmission used the LoRa wireless protocol. Each node was equipped with a solar power module, enabling continuous operation for 21 days even under continuous rainy weather.

[0041] The chemical analysis unit is deployed using a mobile inspection model, equipped with a vehicle-mounted laboratory platform integrating an automated pH analyzer, an organic matter near-infrared spectroscopy detection module, and an ion-selective electrode array. The sampling point layout overlaps with the physical monitoring nodes, but a rotating inspection mechanism is employed, with 12 nodes inspected daily in three batches: pH measurement uses an in-situ inserted composite electrode, calibrated at two points using pH 6.86 and 9.18 standard buffer solutions before measurement; organic matter content detection involves collecting a 5 cm layer of topsoil sample, air-drying and grinding it before placing it in the spectrometer sample chamber, and calculating organic carbon content using partial least squares; nutrient element concentration detection uses vacuum dialysis to extract soil solution, with an ammonium ion electrode detecting nitrogen content, a platinum electrode detecting phosphorus content, and a flame photometer detecting potassium concentration. All chemical test results are immediately uploaded to a central database upon completion, along with the sampling depth, detection time, and instrument status code. The biosensor detection system is designed to combine fixed-point deployment with mobile detection. Fixed biosensors are installed at 9 key points selected from 36 nodes. These sensors utilize microfluidic chip technology, with each chip containing a microbial respiration intensity detection chamber and an enzyme activity reaction chamber. Microbial activity was detected by measuring the carbon dioxide release rate. Soil samples were mixed with nutrient substrate and injected into the reaction chamber, where an infrared gas analyzer recorded CO2 concentration changes every 60 seconds. Enzyme activity was detected targeting urease and phosphatase. Urease was detected using a phenol-sodium hypochlorite colorimetric method to determine ammonia nitrogen release, while phosphatase was detected using the p-nitrophenol method to determine p-nitrophenol production. The mobile detection unit toured all nodes every 72 hours, using a portable microbial respiration meter for on-site detection, and the results were cross-validated with data from fixed sensors.

[0042] The integration of multi-source soil data establishes a unified spatiotemporal framework. After receiving three data streams, the central server first performs time alignment, interpolating data from different acquisition frequencies to a unified 15-minute time granularity. Physical indicators are interpolated linearly, chemical indicators are interpolated using spline interpolation, and biological indicators retain their original measurements. Spatial registration employs Kriging interpolation, generating 10m × 10m grid data from discrete point data. Each grid point contains a complete dataset of physical, chemical, and biological indicators. Data quality control implements an automatic labeling system. Outlier detection for physical data uses the 3σ criterion, chemical data undergoes inter-laboratory comparison, and biological data includes positive control samples for verification. The final multi-source soil data package includes timestamps, geographic coordinates, values ​​for 36 indicators, data quality identifiers, and acquisition device status information.

[0043] The calculation of the health urgency factor incorporates a multi-dimensional assessment model. Time urgency is calculated using a reciprocal function model; when the ratio of remaining time to total time is less than 0.2, the time urgency coefficient increases exponentially. For example, for a project with a standard assessment period of 72 hours, the time urgency coefficient reaches 1.8 with 14 hours remaining. The quantification of health importance employs the analytic hierarchy process (AHP), constructing a three-order judgment matrix for crop type, growth stage, and soil type. The importance weight for corn's tasseling stage is set at 2.3, for wheat's jointing stage at 1.8, and for soybean's pod-setting stage at 2.1. A dynamic adjustment strategy is used for weight coefficient configuration, considering meteorological factors. When extreme weather events are forecast, the system automatically increases the weight of relevant indicators by 0.2. The synthesis of the health urgency factor uses a weighted geometric mean method to avoid a single dimension dominating the calculation results while preserving the differentiated influence of each dimension. For sensor network maintenance, a preventative maintenance system is established. Physical sensors undergo weekly on-site calibration, humidity probes are calibrated using gravity sampling, temperature sensors are compared with standard platinum resistance thermometers, and density meters are verified using a sampling ring knife method. Chemical analysis instruments undergo rigorous quality control procedures. pH electrodes are calibrated before each use, organic matter spectrometers are calibrated with standard samples every 48 hours, and ion electrodes have their sensitive membranes replaced periodically. Biosensors perform daily self-tests, microbial detection chambers run blank controls for each batch, and enzyme activity chips require substrate replacement after every 24 uses. All maintenance records are uploaded to the equipment management system, generating sensor health indices for data reliability assessment.

[0044] The data fusion process employs an adaptive weighting algorithm. Physical data has a base weight of 0.35, but the weight of humidity data increases to 0.45 when rainfall exceeds 10 mm. Chemical data has a base weight of 0.45, but the weight of nutrient element data decreases by 0.1 within 24 hours after fertilization. Biological data has a base weight of 0.2, which increases to 0.3 when abnormal microbial activity is detected. Weight adjustments are determined based on historical data analysis, and the optimal weight combination for each scenario is continuously optimized using machine learning algorithms. Data time consistency checks utilize a sliding window comparison mechanism, initiating verification checks for abruptly changed indicator values ​​to ensure the authenticity of data changes. The application of the health urgency factor reflects priority scheduling: high-urgency tasks (factor > 2.0) are allocated double computing resources, and a fast clustering algorithm is used to shorten processing time; medium-urgency tasks (1.0-2.0) use the standard process; low-urgency tasks (< 1.0) are allowed delayed processing. Task queue management employs priority preemptive scheduling, allowing high-urgency requests to interrupt the computation of low-priority tasks. The system monitoring interface displays the urgency factor and estimated completion time of each task in real time, providing decision support for managers. A full-chain quality traceability system is established throughout the implementation process, recording detailed operation logs and quality indicators at each stage, from sensor data collection to final evaluation results. An automatic review mechanism is triggered for data anomalies; when data anomalies occur for three consecutive cycles at the same node, the system dispatches a drone for on-site investigation.

[0045] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0046] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A soil health assessment method based on multi-source data fusion, characterized in that, Includes the following steps: Multi-source soil data is collected, and soil physical, chemical and biological indicators are acquired in real time through sensors and monitoring equipment. A health indicator matrix is ​​constructed based on the multi-source soil data and normalized to obtain initial health matrix data. Receive a soil health assessment request, parse the soil health assessment request to generate a health description vector, and calculate the health urgency factor to obtain health demand data; Based on the improved clustering algorithm, coarse-grained matching is performed by traversing the initial health matrix data according to the health description vector to obtain a subset of candidate health status data. The candidate health state subset data is fine-grained by using an adaptive neural network model, and the state that maximizes the health reward is selected through a reward mechanism to output the optimal health state and obtain the health assessment instruction. Soil health is assessed according to the health assessment instructions, and real-time data of the optimal health status is monitored. When the real-time data of the optimal health status exceeds a preset threshold, a health data calibration protocol is initiated.

2. The soil health assessment method based on multi-source data fusion as described in claim 1, characterized in that, The collected multi-source soil data includes: Soil physical, chemical, and biological indicators were collected to obtain multi-source soil data. The physical indicators included soil moisture, soil temperature, and soil density. The chemical indicators included pH value, organic matter content, and nutrient element concentration. The biological indicators included microbial activity and enzyme activity. Based on the multidimensional indicators of each monitoring point in the multi-source soil data, a health indicator matrix is ​​established in the order of the monitoring points. The health indicators are normalized using a standard normalization method, so that the indicators of each dimension are within the same range of values, thus obtaining the initial health matrix data.

3. The soil health assessment method based on multi-source data fusion as described in claim 1, characterized in that, The receipt of soil health assessment requests includes: Receive a soil health assessment request, obtain the assessment parameters, health standards and time requirements in the soil health assessment request, and obtain the health assessment parameters; The health assessment parameters are integrated into a health description vector. The current time and the time requirement in the health description vector are obtained. The health urgency factor is calculated based on the pre-set weight coefficients and health level.

4. The soil health assessment method based on multi-source data fusion as described in claim 1, characterized in that, The coarse-grained matching based on the improved clustering algorithm includes: Based on the assessment parameters and health standards in the health description vector, the initial health matrix data is traversed to select monitoring points that meet the basic requirements of health assessment, and the monitoring points are combined into a first candidate health status subset. The first candidate health status subset is sorted from high to low according to the health index scores of the monitoring points to obtain the second candidate health status subset; A health status group is established based on an improved clustering algorithm, with the left side being the health description vector and the right side being the second candidate health status subset; An adjustment variable is introduced to address multiple health constraints, and an optimization function is established. By using group center and boundary search, the top candidate health status is output, and a subset of candidate health status data is obtained.

5. The soil health assessment method based on multi-source data fusion as described in claim 1, characterized in that, The fine-grained optimization of the candidate health state subset data using an adaptive neural network model includes: Define the state space, action space, and reward function of the adaptive neural network model; The real-time state of the candidate health state subset data is used as the input as the state space of the model. The action space is used to generate binary actions of selection and non-selection for each state of the candidate health state subset, and outputs a single optimal health state. The reward function is determined to include a multi-dimensional dynamic reward mechanism, which includes at least positive and negative rewards.

6. The soil health assessment method based on multi-source data fusion as described in claim 5, characterized in that, The state that maximizes health rewards through the reward mechanism includes: Based on the subset of candidate health status data, status data of candidate health status is collected, and actions are selected through an exploration strategy. The exploration strategy includes at least two approaches: probabilistically exploring insufficiently evaluated states and probabilistically selecting the state with the highest current reward value. Calculate the immediate reward based on the actual results, update the reward value using a value update formula, output the optimal health status, and obtain the health assessment instruction.

7. The soil health assessment method based on multi-source data fusion as described in claim 1, characterized in that, The real-time data for monitoring the optimal health status includes: Combine the optimal health status identifier with health assessment parameters into an assessment instruction; Use a health assessor to verify the current availability of health data; If the verification passes, the health assessment process will be initiated; if the verification fails, the data anomaly handling process will be triggered. The system has multiple preset threshold levels, including warning thresholds and emergency thresholds. The warning threshold is defined as soil moisture deviation exceeding a first preset deviation tolerance and pH value deviation exceeding a second preset deviation tolerance for a duration of a first preset duration. The emergency threshold is defined as soil moisture deviation exceeding a third preset deviation tolerance and pH value deviation exceeding a fourth preset deviation tolerance for a duration of a second preset duration. The third preset deviation tolerance is greater than the first preset deviation tolerance, the fourth preset deviation tolerance is greater than the second preset deviation tolerance, and the second preset duration is shorter than the first preset duration. When the real-time data of the optimal health status exceeds the preset threshold of any level, the health data calibration protocol is activated.

8. The soil health assessment method based on multi-source data fusion as described in claim 1, characterized in that, The initiation health data calibration protocol includes: The calibration amount is calculated based on the deviation between real-time health data and standard health data. Based on the calibration values, adjust the health assessment parameters and regenerate the health description vector; The optimal health status is updated by performing coarse-grained matching and fine-grained optimization using the regenerated health description vector.

9. A soil health assessment method based on multi-source data fusion as described in claim 2, characterized in that, The collected soil physical, chemical, and biological data include: Using a soil sensor network to continuously monitor physical parameters, including soil moisture, soil temperature, and soil density; Use a chemical analyzer to regularly measure chemical indicators, including pH value, organic matter content, and nutrient concentration; Use biosensors to detect biological indicators, including microbial activity and enzyme activity; Integrate all data to generate multi-source soil data.

10. The soil health assessment method based on multi-source data fusion as described in claim 3, characterized in that, The calculation of the health urgency factor includes: Based on the time requirements and health standards in the health assessment parameters, and combined with the current time, the time urgency is calculated; Health importance is calculated based on health level and weighting coefficient. By combining time urgency with health importance, a health urgency factor is generated.

Images