A mineral fertilizer intelligent dispensing control method for water and fertilizer integration

Abstract

The application relates to the technical field of intelligent water and fertilizer allocation, and discloses a mineral fertilizer intelligent allocation control method for water and fertilizer integration. The method comprises the following steps: collecting real-time data of soil humidity, pH value and nutrient concentration through a wireless sensor network, and synchronously acquiring precipitation, air temperature and light intensity data of a remote meteorological server; combining environmental data and a crop growth calendar to construct a soil nutrient consumption simulation model; after being corrected according to real-time monitoring data, generating a farmland nutrient deficiency distribution map by using a geographic statistics method; inputting the map into a fertilizer recommendation algorithm to calculate the application amount per unit area, and simultaneously analyzing the trends of the two types of data to identify abnormal fertilization; when the application amount exceeds the safe range, triggering multi-level alarms and adjusting the amount of fertilizer to be put according to the abnormal results. The method is suitable for the dynamic changes of soil and environment, and realizes intelligent and accurate fertilizer allocation.

Description

Technical Field

[0001] This invention relates to the field of intelligent water and fertilizer distribution technology, specifically to an intelligent distribution and control method for mineral fertilizers used in integrated water and fertilizer management. Background Technology

[0002] In the process of agricultural production transforming towards modernization and intensification, integrated water and fertilizer technology has become an important technical means to improve agricultural production efficiency because it enables the synergistic utilization of water resources and fertilizers. Mineral fertilizers, as a crucial source of nutrients for crop growth, have their allocation directly impacting crop growth and the farmland ecological environment. Currently, traditional mineral fertilizer allocation methods largely rely on growers' experience and judgment, lacking dynamic awareness of farmland soil and environmental conditions. Allocation schemes are often uniform and outdated, making it difficult to adapt to the heterogeneity of soils in different regions and the dynamic changes in environmental factors.

[0003] Soil moisture, pH, and nutrient concentration are key soil indicators affecting fertilizer absorption efficiency. Different plots of land exhibit natural variations in soil characteristics, and traditional fertilizer application methods fail to accurately monitor these indicators in real time, leading to a mismatch between fertilizer application rates and actual soil needs. While some existing technologies incorporate sensors to collect soil data, they fail to fully consider the impact of meteorological factors on soil nutrient consumption. Changes in precipitation, temperature, and light intensity directly alter the rate of soil nutrient decomposition and crop absorption efficiency; ignoring this environmental observation data results in insufficient predictive accuracy in fertilizer application models.

[0004] Existing soil nutrient consumption models mostly use fixed parameter settings and fail to dynamically correct based on actual farmland monitoring data. This leads to discrepancies between model predictions and actual nutrient consumption, affecting the accuracy of fertilizer application. Furthermore, nutrient consumption rates vary across different areas of farmland. Traditional application methods do not provide precise regional assessments of farmland, making it difficult to generate targeted nutrient deficiency distribution data, resulting in localized over- or under-application of fertilizer.

[0005] Excessive fertilization not only wastes mineral fertilizers and increases agricultural production costs, but may also lead to ecological and environmental problems such as soil compaction and eutrophication of water bodies. Insufficient fertilization, on the other hand, results in nutrient deficiencies in crops, affecting yield and quality. More critically, existing technologies lack effective mechanisms for identifying abnormal events during fertilization, failing to promptly detect instances where the fertilization amount exceeds safe limits, and hindering rapid adjustments, further exacerbating the negative impacts of irrational fertilizer application. Furthermore, while some technologies include simple alarm functions, they lack adjustment mechanisms linked to anomaly detection results, requiring manual intervention after alarms, resulting in low response efficiency and failing to meet the demands of modern agriculture for real-time and intelligent fertilizer application. With the continuous improvement of agricultural intelligence, a smart mineral fertilizer application and control method is needed that can integrate dynamic soil and environmental data to achieve precise application and anomaly control. Summary of the Invention

[0006] The purpose of this invention is to provide an intelligent distribution and control method for mineral fertilizers in fertigation to solve the problems mentioned in the background art.

[0007] To achieve the above objectives, the present invention provides a method for intelligent dispensing and control of mineral fertilizers for fertigation, the method comprising:

[0008] Real-time monitoring data on soil moisture, pH, and nutrient concentration are collected through a wireless sensor network set up in farmland; at the same time, environmental observation data on precipitation, temperature, and light intensity are obtained from a remote meteorological server.

[0009] A soil nutrient consumption simulation model was constructed using the environmental observation data and crop growth calendar, and the parameters of the soil nutrient consumption simulation model were corrected using the real-time monitoring data.

[0010] Based on the corrected soil nutrient consumption simulation model, a nutrient deficiency distribution map of farmland area was generated using geostatistical methods.

[0011] The nutrient deficiency distribution map is input into the fertilizer recommendation algorithm to calculate the application rate of mineral fertilizer per unit area. The algorithm also performs trend consistency analysis on the real-time monitoring data and the environmental observation data to identify abnormal fertilization events. When the application rate per unit area exceeds the safe range, a multi-level alarm protocol is triggered and the fertilizer application rate is adjusted based on the identification results of the abnormal fertilization events.

[0012] Preferably, collecting real-time monitoring data on soil moisture, pH, and nutrient concentration through a wireless sensor network set up in the farmland includes: deploying multiple sensor nodes in each irrigation zone of the farmland, with each sensor node measuring soil parameters at preset time intervals, and aggregating the measurement data into a time-series dataset through a gateway device as real-time monitoring data.

[0013] Preferably, constructing a soil nutrient consumption simulation model using the environmental observation data and crop growth calendar includes: determining the nutrient absorption pattern at different growth stages based on the crop growth calendar, calculating the natural loss rate of soil nutrients by combining precipitation and temperature data; integrating the nutrient absorption pattern with the natural loss rate to establish a nutrient balance equation based on a time step; and training the nutrient balance equation using historical soil data to form a soil nutrient consumption simulation model.

[0014] Preferably, the method of generating a nutrient deficiency distribution map of farmland area based on the corrected soil nutrient consumption simulation model and applying geostatistical methods includes: extracting spatial coordinates and nutrient values ​​from the real-time monitoring data, calculating a semi-variogram to describe spatial correlation; estimating nutrient levels at unsampled locations using the Kriging interpolation algorithm to generate an initial nutrient distribution surface; comparing the initial nutrient distribution surface with the output of the soil nutrient consumption simulation model, and generating a nutrient deficiency distribution map through iterative optimization.

[0015] Preferably, the nutrient deficiency distribution map is input into a fertilizer recommendation algorithm to calculate the application rate of mineral fertilizer per unit area, and a trend consistency analysis is performed on the real-time monitoring data and the environmental observation data to identify abnormal fertilization events, including:

[0016] Nutrient thresholds are set based on crop type and soil characteristics, and the amount of fertilizer applied at the minimum cost is calculated using linear programming.

[0017] A sliding window correlation calculation is performed on the nutrient concentration sequence in the real-time monitoring data and the precipitation sequence in the environmental observation data to generate an anomaly index; when the anomaly index exceeds a dynamic threshold, it is marked as a fertilization anomaly event.

[0018] Preferably, the step of generating the nutrient deficiency distribution map through iterative optimization includes: setting a nutrient deficiency tolerance range; comparing the difference comparison results with the tolerance range, initiating a local densification sampling command for grid areas exceeding the tolerance range, driving the UAV to carry a mobile sensor for supplementary measurement; updating the initial nutrient distribution surface using the supplementary measurement data, and repeating the difference comparison step until the differences in all grid areas fall within the tolerance range, and finally outputting the optimized nutrient deficiency distribution map.

[0019] Preferably, the calculation of the semivariogram to describe spatial correlation includes: fitting the semivariogram with an exponential model to calculate the nugget value, sill value, and range parameter; selecting the optimal semivariogram model using cross-validation and verifying the spatial correlation through spatial autocorrelation analysis; the kriging interpolation algorithm uses ordinary kriging and combines the semivariogram parameters to generate a spatial interpolation weight matrix, thereby estimating the nutrient level at unsampled locations.

[0020] Preferably, the generation of the anomaly index includes: calculating the Pearson correlation coefficient and Spearman rank correlation coefficient of the nutrient concentration sequence and the precipitation sequence within a sliding window; weighting and fusing the Pearson correlation coefficient and Spearman rank correlation coefficient, and introducing a baseline offset based on historical normal data to jointly constitute the anomaly index; the dynamic threshold is dynamically determined based on the anomaly index sequence of a preset period by calculating its mean and standard deviation, and using the three sigma criterion.

[0021] Preferably, training the nutrient balance equation using historical soil data includes: collecting historical soil nutrient data, environmental observation data, and crop yield data for at least one growing season; constructing a deep neural network model, using the nutrient balance equation as the initial model, inputting historical data into the network for supervised learning, and optimizing the model parameters through a backpropagation algorithm; and using a validation set to cross-validate the trained model, thereby forming a calibrated soil nutrient consumption simulation model.

[0022] Preferably, the step of triggering a multi-level alarm protocol and adjusting the fertilizer application rate based on the identification results of the fertilization anomaly includes: dividing the alarm levels into three levels: prompt, warning, and severe; when the application rate per unit area exceeds the safe range but is within a preset buffer, a prompt-level alarm is triggered, only logging is recorded and the agronomist is notified; when the application rate continues to exceed the safe range and an fertilization anomaly is identified, a warning-level alarm is triggered, the system automatically reduces the fertilizer application rate to the upper limit of the safe range, and sends an SMS notification; when the application rate severely exceeds the safe range and is accompanied by abnormal data from key sensors, a severe-level alarm is triggered, the system immediately stops fertilizer application, starts the backup irrigation system for flushing with clean water, and simultaneously sends an emergency notification to the agronomist and administrator.

[0023] Compared with the prior art, the beneficial effects of the present invention are:

[0024] Through collaborative data collection via wireless sensor networks and remote meteorological servers, comprehensive perception of core indicators of farmland soil and key environmental factors has been achieved. Real-time monitoring data on soil moisture, pH, and nutrient concentration directly reflect the soil's current nutrient supply capacity and physicochemical state, while environmental observation data such as precipitation, temperature, and light intensity fully present the external conditions affecting soil nutrient decomposition, transformation, and crop absorption. The combination of these two types of data breaks through the limitations of single data sources in traditional allocation methods, providing a more comprehensive and realistic basis for fertilizer allocation schemes.

[0025] The soil nutrient consumption simulation model was constructed by combining environmental observation data with crop growth calendars. It fully considers the differences in nutrient requirements at different crop growth stages and the dynamic impact of environmental factors on nutrient consumption, enabling the model to closely align with crop growth patterns and environmental changes. By using real-time monitoring data to calibrate the model's parameters, deviations between the model and actual farmland conditions can be dynamically corrected, making the model's predictions more consistent with the true state of soil nutrient consumption and providing a more reliable basis for subsequent fertilizer application.

[0026] The application of geostatistical methods can accurately capture the differences in soil nutrient consumption in different areas of farmland. The generated nutrient deficiency distribution map clearly presents the nutrient distribution pattern within the farmland, breaking the limitations of the traditional method of uniform fertilizer application across the entire farmland and providing an intuitive reference for precise regional fertilizer application. Based on this distribution map, the fertilizer recommendation algorithm can calculate the corresponding application rate per unit area according to the nutrient deficiency situation in different areas, making fertilizer application more targeted and avoiding the problem of over- or under-fertilization in local areas.

[0027] Consistency analysis of trends in real-time monitoring data and environmental observation data can promptly capture abnormal fluctuations during fertilization, effectively identify abnormal fertilization events, and quickly detect risk points in fertilizer formulation. The multi-level alarm protocol triggering mechanism can issue timely warnings when the application rate per unit area exceeds the safe range. Adjustments to fertilizer application based on anomaly identification results achieve linkage between alarms and control, enabling rapid response without manual intervention and improving the safety and flexibility of the fertilizer formulation process.

[0028] The entire method forms a complete closed loop from data acquisition, model building, distribution assessment to allocation calculation, anomaly identification, and alarm adjustment, realizing the intelligent operation of mineral fertilizer allocation. Its adaptation to the spatial heterogeneity of farmland soil characteristics and the dynamic changes in soil and environmental conditions allows for real-time optimization of fertilizer allocation schemes. This avoids waste of mineral fertilizers, reduces agricultural production costs, minimizes the negative impact of excessive fertilization on the ecological environment, and ensures that crops receive appropriate nutrient supply, thus contributing to the coordinated development of agricultural production in terms of efficiency, quality, and ecological protection. Attached Figure Description

[0029] Figure 1 This is a schematic diagram illustrating the working principle of the intelligent mineral fertilizer dispensing and control method for fertigation as described in this invention.

[0030] Figure 2 Flowchart for constructing a soil nutrient consumption simulation model;

[0031] Figure 3 A flowchart generated for a nutrient deficiency distribution map;

[0032] Figure 4 A map for dynamic monitoring and anomaly identification of soil nutrients;

[0033] Figure 5 This is a graph showing the abnormal fertilization detection index and dynamic threshold analysis. Detailed Implementation

[0034] 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.

[0035] Please see Figure 1 This invention provides an intelligent distribution and control method for mineral fertilizers in fertigation, which optimizes fertilizer application by integrating real-time monitoring data and environmental observation data. Specific implementation details are as follows:

[0036] A wireless sensor network deployed in the farmland collects real-time monitoring data on soil moisture, pH, and nutrient concentration, while simultaneously acquiring environmental observation data on precipitation, temperature, and light intensity from a remote meteorological server. A soil nutrient consumption simulation model is constructed using environmental observation data and a crop growth calendar, and the model's parameters are calibrated using real-time monitoring data to ensure accuracy. Based on the calibrated model, a nutrient deficiency distribution map of the farmland area is generated using geostatistical methods, visually displaying areas of nutrient deficiency. This nutrient deficiency distribution map is input into a fertilizer recommendation algorithm to calculate the application rate of mineral fertilizer per unit area. Simultaneously, trend consistency analysis is performed on the real-time monitoring data and environmental observation data to identify abnormal fertilization events. When the application rate per unit area exceeds the safe range, the system triggers a multi-level alarm protocol and dynamically adjusts the fertilizer application rate based on the identification results of abnormal fertilization events, thereby achieving precise and safe fertilizer management.

[0037] Example 1: See Figure 2The irrigation zones for farmland are divided based on topographic slope, soil type, and crop planting layout. Within each irrigation zone, multiple sensor nodes are arranged in a grid pattern. The density of sensor nodes is adjusted according to the uniformity of the farmland; the number of sensor nodes is increased in areas with significant variations in soil properties. Each sensor node integrates a soil moisture sensor, a pH sensor, and a nutrient concentration sensor. The nutrient concentration sensor primarily monitors the content of elements such as nitrogen, phosphorus, and potassium. The sensor nodes perform measurements at preset time intervals, which can be flexibly configured according to different crop growth stages. For example, during the vigorous growth period, the preset time interval is set to once every hour, while during the dormant period, it is set to once every four hours. After measurement, each sensor node transmits the collected soil moisture, pH, and nutrient concentration data to the regional gateway device via low-power wireless communication protocols such as ZigBee or LoRa.

[0038] In practice, the regional gateway device receives data from all sensor nodes within its coverage area. The regional gateway performs preliminary verification and packaging of the received data, aggregating data from different sensor nodes at the same timestamp to form a structured time-series dataset, which constitutes the real-time monitoring data. The regional gateway device uploads the aggregated real-time monitoring data to a central data processing platform in the cloud via 4G or 5G networks. Simultaneously, the central data processing platform establishes a connection with a remote meteorological server through a pre-defined application programming interface (API) to periodically acquire precipitation, temperature, and light intensity data for the farmland area. These data collectively constitute the environmental observation data. The environmental observation data and real-time monitoring data are aligned and stored according to timestamps in the central data processing platform, providing a data foundation for subsequent modeling.

[0039] This process involves constructing a soil nutrient consumption simulation model using environmental observation data and a crop growth calendar. The crop growth calendar defines in detail the time range of each physiological stage of a specific crop from sowing, emergence, growth, flowering to maturity, as well as the typical nutrient requirements of each stage. Based on the crop growth calendar, nutrient absorption patterns at different growth stages can be determined. A nutrient absorption pattern is a function describing the changes in the amount of various nutrients absorbed by the crop per unit time. Combining precipitation and temperature data obtained from a remote meteorological server, the natural nutrient loss rate of the soil can be calculated. The natural loss rate mainly considers nutrient decomposition or volatilization caused by rainfall leaching and temperature effects. By integrating the nutrient absorption pattern with the natural loss rate, a time-step-based nutrient balance equation is established. This equation uses days as the basic time step to simulate the net change in soil nutrients each day. Its core principle is that the input term minus the output term equals the change in soil nutrient inventory.

[0040] In practical implementation, training the nutrient balance equation using historical soil data is a crucial step. This requires collecting historical data from at least one complete growing season, including but not limited to historical soil nutrient data, historical environmental observation data, and final crop yield data. A deep neural network model is constructed, using the aforementioned nutrient balance equation as the initial model structure. Historical soil nutrient data and historical environmental observation data are used as input features, and crop yield data or measured soil nutrient data from key growing seasons are used as training labels. These are then input into the deep neural network model for supervised learning. The training process employs the backpropagation algorithm, calculating the loss function between predicted values ​​and true labels to iteratively optimize the weights and bias parameters in the deep neural network model, gradually approximating the actual situation in the model's output. After training, the dataset is divided into training and validation sets. The validation set is used to cross-validate the trained deep neural network model, evaluating its generalization ability and preventing overfitting. The final result is a calibrated soil nutrient consumption simulation model that accurately reflects the nutrient dynamics of a specific field.

[0041] It is understandable that the stable operation of the sensor network is fundamental to data quality. In some embodiments, sensor nodes are powered by a combination of solar panels and rechargeable batteries to ensure energy needs for long-term field operations. Sensor nodes have built-in self-diagnostic functions, capable of detecting the working status of sensor probes and sending status alarm information to the area gateway device when data anomalies or hardware failures are detected. In addition to data aggregation and forwarding, the area gateway device also possesses edge computing capabilities, enabling it to perform preliminary filtering and outlier removal on real-time monitoring data, reducing the upload of invalid data.

[0042] It is understandable that the accuracy of model construction depends on the completeness of the data. In some embodiments, for environmental observation data obtained from remote meteorological servers, if short-term data gaps occur, the central data processing platform will use time-series interpolation methods to fill in the gaps, such as linear interpolation or spatial interpolation methods based on data from nearby meteorological stations. The crop growth calendar is not fixed but can be dynamically fine-tuned based on local phenological observation data to make nutrient absorption patterns more consistent with actual crop growth conditions. The structure of the deep neural network model can be selected from multilayer perceptrons or long short-term memory networks to capture the nonlinear relationships and temporal dependencies of nutrient consumption.

[0043] Optionally, historical data collection can be extended to multiple growing seasons, encompassing data from different climatic years. This helps improve the robustness of soil nutrient consumption simulation models under abnormal climatic conditions. During model training, regularization techniques can be introduced to further optimize training performance and prevent deep neural network models from over-relying on certain specific features in the training data. The calibrated soil nutrient consumption simulation model will periodically utilize the latest real-time monitoring data for incremental learning, enabling continuous optimization and adaptive updates.

[0044] Optionally, the nutrient balance equation can be established by considering more complex factors. For example, when calculating the natural loss rate, in addition to precipitation and temperature, light intensity data can be incorporated, as light affects soil microbial activity and thus indirectly influences nutrient transformation. The output of the soil nutrient consumption simulation model can not only be used to predict nutrient deficiency but also, in conjunction with crop models, predict potential yields, providing more comprehensive decision support for farm management. The entire process of data acquisition, transmission, processing, and control command issuance forms a closed-loop system, realizing the automation and intelligence of integrated water and fertilizer management.

[0045] Example 2: See Figure 3 To generate a nutrient deficiency distribution map based on a corrected soil nutrient consumption simulation model, the first step is to process real-time monitoring data from a wireless sensor network. This data includes the geographic coordinates of each sensor node and its measured nutrient concentration. The initial step in geostatistical methods is to extract these spatial coordinates and corresponding nutrient values, forming a sample point set for spatial analysis. A semivariogram is calculated to describe the spatial correlation between these sample points. The semivariogram is a fundamental tool in geostatistics for quantifying spatial autocorrelation, constructed by calculating the variance between sample point pairs at different distance intervals. An exponential model is used to fit the calculated empirical semivariogram. The exponential model is one of the commonly used theoretical variogram models. The fitting process involves calculating three key parameters: nugget value, sill value, and range parameter. The nugget value represents the random variation at the minimum sampling distance, the sill value represents the value when the semivariogram reaches stationarity, and the range parameter describes the maximum distance range where spatial autocorrelation exists. To ensure the accuracy of the semivariogram model, cross-validation is required to select the optimal semivariogram model from various candidate models such as the exponential model, the spherical model, and the Gaussian model. The significance of spatial correlation is then verified using spatial autocorrelation analysis indicators such as the Moran index.

[0046] In practical implementation, estimating the nutrient levels at unsampled locations using the Kriging interpolation algorithm is the core of generating the initial nutrient distribution surface. The Kriging interpolation algorithm is an optimal unbiased estimation method based on semi-variogram theory. This method specifically employs ordinary Kriging, which assumes that the expected value of the variables within the study area is an unknown constant. The ordinary Kriging method, combined with the semi-variogram parameters obtained in previous steps, generates a spatial interpolation weight matrix. Each weight value in the spatial interpolation weight matrix represents the contribution of known sample points to the estimated value of unknown points. The determination of the weight values ​​follows the conditions of minimizing the estimation variance and being unbiased. This spatial interpolation weight matrix is ​​applied to the weighted sum of the nutrient values ​​of all known sample points, thereby calculating the estimated nutrient level at each regular grid point within the farmland area. Connecting these estimates forms the initial nutrient distribution surface covering the entire farmland area. The initial nutrient distribution surface is a continuous digital surface model that reflects the spatial distribution of nutrients based on spatial interpolation.

[0047] The initial nutrient distribution surface is compared with the output of the soil nutrient consumption simulation model, which outputs simulated nutrient levels for the farmland area at the same time point. The difference comparison is performed on each identical grid cell, calculating the difference between the spatial interpolation estimate provided by the initial nutrient distribution surface and the dynamic simulation value provided by the soil nutrient consumption simulation model. A final nutrient deficiency distribution map is generated through iterative optimization, and a nutrient deficiency tolerance range is defined, specifying an acceptable range of difference based on agronomical knowledge. The difference comparison results for each grid cell are compared with the nutrient deficiency tolerance range to identify grid areas where the difference exceeds the tolerance range. For these grid areas exceeding the tolerance range, the system initiates a local densification sampling command, which drives a UAV carrying a mobile sensor to fly to the designated area for supplementary measurements. The UAV performs a low-altitude, refined scan of the target grid area to acquire supplementary measurement data. This newly acquired supplementary measurement data is used to update the previous sample point set, and the semi-variogram function and Kriging interpolation are recalculated to generate an updated initial nutrient distribution surface. Then, the difference comparison step is executed again, comparing the updated initial nutrient distribution surface with the output of the soil nutrient consumption simulation model. This iterative process is repeated until the difference comparison results for all grid regions fall within the preset nutrient deficiency tolerance range. At this point, the iterative optimization process terminates, and the system outputs the final optimized nutrient deficiency distribution map. The optimized nutrient deficiency distribution map combines the advantages of real-time spatial measurement data and process simulation data, resulting in higher accuracy.

[0048] It is understandable that the computational quality of the semivariogram directly affects the accuracy of spatial interpolation. In some embodiments, before calculating the semivariogram, it is necessary to preprocess the nutrient values ​​in the real-time monitoring data, such as checking and processing possible outliers or obvious erroneous data caused by sensor malfunctions, to avoid these noisy data having an excessive impact on the structure of the semivariogram. When fitting the exponential model, optimization algorithms such as the least squares method can be used to solve for the optimal solutions of nugget value, sill value, and range parameter, so that the theoretical model can best fit the empirically calculated semivariogram points. The cross-validation method usually adopts the leave-one-out method, that is, one sample point is excluded in turn, and the value of the excluded point is estimated by the current semivariogram model and Kriging method using the remaining sample points. Then, the error between the estimated value and the true value is compared, and the model with the smallest average error is selected as the optimal semivariogram model.

[0049] It is understandable that iterative optimization is key to balancing computational cost and accuracy. In some embodiments, the nutrient deficiency tolerance range can be dynamically adjusted based on the crop's sensitivity to nutrients. For crops in critical growth stages or sensitive to nutrient stress, the nutrient deficiency tolerance range is set narrower to achieve higher accuracy; while for crops in non-critical periods or tolerant of poor soil, the nutrient deficiency tolerance range can be appropriately widened to reduce the number of UAV flights and the consumption of computational resources. The UAV's flight path planning algorithm comprehensively considers the spatial distribution of the grid areas requiring intensive sampling and automatically generates a flight path with optimal energy efficiency to cover all areas requiring supplementary measurements in the shortest possible flight time. The mobile sensors used for supplementary measurements need to be calibrated regularly to ensure that their measurement results are consistent and comparable with the data from the fixed sensor network deployed in the farmland.

[0050] Optionally, the implementation of the kriging interpolation algorithm can consider more complex situations. For example, when there are obvious trends in farmland areas, a generalized kriging method can be used, which can simultaneously handle spatial correlation and deterministic trends. When generating the initial nutrient distribution surface, in addition to the nutrient concentration itself, auxiliary variables such as elevation and slope can be introduced for co-kriging interpolation, which may improve interpolation accuracy in areas with sparse sample points. The convergence condition during the iterative optimization process can be set with a maximum number of iterations to prevent infinite loops due to data issues in individual areas. When the maximum number of iterations is reached and some areas still do not meet the tolerance conditions, the system will issue a prompt message, suggesting manual verification.

[0051] Optionally, the output format of the nutrient deficiency distribution map can be a raster data file, where each raster cell value represents the degree of nutrient deficiency at that location. This format facilitates further visualization and spatial analysis by the geographic information system. The degree of nutrient deficiency can be expressed as an absolute value or as a relative percentage relative to the crop's requirement threshold. The generated nutrient deficiency distribution map can be overlaid with farmland boundary maps, soil type maps, etc., providing agronomists with a more comprehensive spatial decision-making context. The entire process of generating the nutrient deficiency distribution map can be set to run automatically on a regular basis, such as weekly, to dynamically track the spatiotemporal changes in nutrient deficiency during the crop growing season.

[0052] Example 3: A nutrient deficiency distribution map is input into a fertilizer recommendation algorithm to calculate the application rate of mineral fertilizers per unit area. The algorithm first sets nutrient thresholds based on crop type and soil characteristics. These thresholds represent the minimum required concentration and maximum safe concentration of each nutrient element, determined based on crop nutrition knowledge and soil fertility standards. Linear programming is then used to calculate the fertilizer application rate at the minimum cost. The linear programming method uses minimizing fertilizer purchase cost as its objective function, which is subject to various constraints, including but not limited to: the application rate per unit area must meet the nutrient threshold requirements; the fertilizer mixing ratio cannot exceed the safe range; and the total fertilizer application cannot exceed the carrying capacity of the farmland. The decision variables in the linear programming problem are the application rates of various mineral fertilizers. The linear programming problem is solved using the simplex method or interior point method to obtain an economical and efficient fertilizer ratio and application rate scheme. Simultaneously, trend consistency analysis is performed on real-time monitoring data and environmental observation data to identify fertilization anomalies. The trend consistency analysis is conducted on the nutrient concentration sequence in the real-time monitoring data and the precipitation sequence in the environmental observation data. The nutrient concentration sequence is the soil nutrient measurement value recorded by the sensor network in time sequence, and the precipitation sequence is the rainfall data provided by the meteorological server recorded in time sequence.

[0053] In practice, a sliding window correlation calculation is performed on the nutrient concentration and precipitation series to generate an anomaly index. The sliding window correlation calculation divides the time series data into continuous time windows of fixed length, calculating the statistical correlation between the two series within each window. The Pearson correlation coefficient and Spearman rank correlation coefficient are calculated separately for the nutrient concentration and precipitation series within the sliding window. The Pearson correlation coefficient measures the linear correlation between the two series, while the Spearman rank correlation coefficient measures the monotonic correlation. The Pearson and Spearman rank correlation coefficients are then weighted and fused. Different weighting coefficients are assigned to the Pearson and Spearman rank correlation coefficients, determined based on historical data analysis, reflecting the relative importance of the two correlation coefficients in anomaly detection. A baseline offset based on historical normal data is introduced. This baseline offset is the average correlation coefficient calculated from periods without fertilization anomalies, and together they constitute the anomaly index. The formula for calculating the anomaly index is:

[0054]

[0055] in, Indicates an abnormality index. This represents the Pearson correlation coefficient. Represents the Spearman rank correlation coefficient. The weights of the Pearson correlation coefficient are indicated. The weights represent the Spearman rank correlation coefficients. Indicates the baseline offset. Weights and It is a positive real number and satisfies baseline offset It is a correction term calculated based on the mean of the correlation coefficient sequence during historical normal periods.

[0056] In practice, the dynamic threshold is dynamically determined based on an anomaly index sequence over a preset period. This preset period is typically set to a complete crop growing season or several weeks. The threshold is dynamically determined by calculating the mean and standard deviation of the anomaly index sequence and applying the three-sigma criterion. Specifically, the process involves collecting the anomaly index values ​​calculated from all sliding windows within the preset period to form an anomaly index sequence, calculating the arithmetic mean and standard deviation of this sequence, and setting the dynamic threshold as the mean plus three times the standard deviation. When the real-time calculated anomaly index exceeds the dynamic threshold, the system marks that time window as a fertilization anomaly event. During the generation of the nutrient deficiency distribution map, accuracy is ensured through iterative optimization. A nutrient deficiency tolerance range is set, defining the maximum acceptable deviation between the estimated and simulated nutrient values. The difference comparison results are compared with the tolerance range. The difference comparison results are the difference between the initial nutrient distribution surface and the output value of the soil nutrient consumption simulation model. Local densification sampling is initiated for grids that exceed the tolerance range. The local densification sampling command drives the UAV to carry the mobile sensor to perform supplementary measurements. The supplementary measurement data is used to update the initial nutrient distribution surface. The difference comparison steps are repeated until the differences in all grid areas fall within the tolerance range.

[0057] It is understandable that the implementation of fertilizer recommendation algorithms relies on accurate nutrient threshold settings. In some embodiments, the nutrient threshold can be fine-tuned based on the genetic characteristics of the crop variety and local soil background values. For example, for high-yield crop varieties, the nutrient threshold may be set higher to ensure sufficient supply, while for sensitive crops, the nutrient threshold may be set lower to prevent fertilizer damage. Constraints in linear programming methods can include environmental constraints, such as maximum nitrogen and phosphorus loss, to comply with sustainable agriculture guidelines. When solving linear programming problems, commercial optimization software libraries or custom algorithms can be used to ensure computational efficiency and the feasibility of the solution.

[0058] It is understandable that the effectiveness of trend consistency analysis depends on the choice of sliding window size. In some embodiments, the length of the sliding window can be dynamically adjusted according to the crop growth stage; the window length is shorter during rapid growth to capture rapid changes, and longer during stable growth to smooth random fluctuations. The calculation of Pearson correlation coefficient and Spearman rank correlation coefficient requires the sequence data to meet certain statistical assumptions. For example, the Pearson correlation coefficient requires the data to be approximately normally distributed. Therefore, before calculation, it may be necessary to transform the nutrient concentration sequence and precipitation sequence, such as through logarithmic transformation, to improve the distribution pattern. The weighting coefficients in the weighted fusion can be optimized using machine learning methods, such as training a logistic regression model using historical outlier data, to determine the optimal weights for the Pearson correlation coefficient and Spearman rank correlation coefficient.

[0059] Understandably, the adaptive capability of dynamic thresholds is important for handling seasonal variations. In some implementations, the three-sigma criterion can be replaced with other statistical rules, such as using moving average control charts. The dynamic threshold is updated based on local statistics of the recent anomalous index sequence, rather than global statistics, to improve the responsiveness to trend changes. After marking fertilization anomalies, the system can record the time, location, and anomalous index value of the event, forming an anomaly event log for subsequent analysis and auditing.

[0060] Optionally, the iterative optimization process can be combined with a real-time control loop. Upon detecting an abnormal fertilization event, the system can not only adjust the fertilizer application rate but also trigger the regeneration of the nutrient deficiency distribution map, thus forming a closed-loop feedback. When the UAV performs localized intensive sampling, a multispectral camera can be integrated for remote sensing measurements to supplement the insufficient point sensor data and improve the accuracy of spatial interpolation. The nutrient deficiency tolerance range can be set to an asymmetric range, for example, the tolerance for nutrient deficiency is lower than the tolerance for nutrient excess, in order to prioritize the prevention of nutrient deficiency symptoms.

[0061] See Figure 4 This paper presents the comprehensive analysis results of soil nutrient dynamic monitoring and anomaly event identification. The charts show the concentration trends of the three main nutrient elements—nitrogen, phosphorus, and potassium—in time series format, overlaid with the distribution of precipitation. The red curve clearly shows the fluctuation pattern of nitrogen throughout the monitoring period, the blue curve reflects the changes in phosphorus content, and the green curve shows the dynamic characteristics of potassium. This nutrient concentration data comes from real-time monitoring by a wireless sensor network deployed in farmland. The cyan bar chart represents precipitation data obtained from a meteorological server, providing environmental observational evidence for trend consistency analysis. The vertical orange dashed lines in the charts mark fertilization anomalies detected by the system. These anomalies are identified based on a sliding window correlation calculation method. When a significant anomaly occurs in the statistical correlation between the nutrient concentration series and the precipitation series, the system triggers an anomaly alarm mechanism. The anomaly detection algorithm comprehensively utilizes the Pearson correlation coefficient and the Spearman rank correlation coefficient, generating an anomaly index through weighted fusion, and correcting it by combining it with baseline offset calculated from historical normal data. Accurate identification of these abnormal events is of great significance for timely adjustment of fertilizer application strategies, prevention of nutrient loss and environmental pollution, and demonstrates the practical application value of intelligent fertilization systems in precision agriculture.

[0062] Example 4: The process of generating the anomaly index involves in-depth analysis of time series data. The anomaly index is calculated by performing a sliding window correlation calculation on the nutrient concentration series in real-time monitoring data and the precipitation series in environmental observation data. The Pearson correlation coefficient and Spearman rank correlation coefficient are calculated for the nutrient concentration series and precipitation series within the sliding window, respectively. The Pearson correlation coefficient quantifies the strength and direction of the linear relationship between the two series, and its calculation depends on the actual numerical value and distribution characteristics of the data. The Spearman rank correlation coefficient focuses on assessing the consistency of the ranking order of the two series, is less sensitive to outliers, and can capture non-linear monotonic relationships. The calculated Pearson and Spearman rank correlation coefficients are then weighted and fused. The weighting and fusion process assigns specific weight coefficients to the Pearson and Spearman rank correlation coefficients. These weight coefficients are fixed values ​​determined based on historical data analysis and reflect the predictive importance of different correlation strengths for fertilization anomaly events under specific farmland conditions. Introducing a baseline offset based on historical normal data is a crucial step in constructing the anomaly index. The baseline offset is the long-term average value of the weighted fusion of Pearson correlation coefficient and Spearman rank correlation coefficient, calculated from historical data during periods not marked as fertilization anomalies. This baseline offset serves as a reference point to correct for background fluctuations in the current calculated value. The Pearson correlation coefficient, Spearman rank correlation coefficient, their respective weighting coefficients, and the baseline offset together constitute the formula for calculating the anomaly index. The magnitude of the anomaly index directly reflects the degree to which the relationship between current nutrient concentration changes and precipitation changes deviates from historical normal patterns.

[0063] The dynamic threshold is dynamically determined based on an anomaly index sequence over a preset period. This preset period is typically chosen to represent a complete agricultural operation phase or a seasonally representative timeframe, such as thirty consecutive days or a complete crop phenological period. The system collects all anomaly index values ​​calculated for each sliding window within this preset period, forming a time-series dataset. The arithmetic mean and standard deviation of this anomaly index sequence are calculated. The arithmetic mean represents the central trend of the anomaly index within that period, while the standard deviation measures the dispersion of the anomaly index around the mean. The three-sigma criterion is used to set the dynamic threshold; specifically, the dynamic threshold is set to the mean of the anomaly index sequence plus three times the standard deviation. When the anomaly index value of the current sliding window, calculated in real-time, exceeds this dynamic threshold, the system determines that a fertilization anomaly event has occurred within that time window and marks and records it accordingly. This dynamic threshold determination method allows the system to adapt to the inherent fluctuations in the data, reducing false alarms caused by seasonal variations or long-term trends.

[0064] In practical implementation, training the nutrient balance equation using historical soil data is the core step in constructing a high-precision soil nutrient consumption simulation model. Historical soil data, historical environmental observation data, and crop yield data for at least one growing season are collected. Historical soil data should include nutrient content measurements at different soil depths and sampling times. Historical environmental observation data should cover meteorological factors during the same period. Crop yield data serves as the final validation indicator for the model's training effectiveness. A deep neural network model is constructed, using the aforementioned physicochemical process-based nutrient balance equation as the initial framework for the deep neural network model. Key parameters involved in the nutrient balance equation, such as the nutrient uptake rate coefficient and the natural loss rate coefficient, are set as parameters to be optimized in the deep neural network model. The collected historical soil data and historical environmental observation data are used as input features to the deep neural network model, while soil nutrient data measured during the same period or the final crop yield data are used as training target labels, inputting them into the deep neural network model for supervised learning. The training process employs the backpropagation algorithm, which calculates the loss function between the predicted values ​​of the deep neural network model and the true labels. Then, iterative adjustments to the weights and bias parameters in the deep neural network model are made using optimization algorithms such as gradient descent, so that the output of the deep neural network model gradually approximates the actual changes in soil nutrients or crop yield results.

[0065] In practice, cross-validating the trained deep neural network model using a validation set is a necessary step to evaluate the model's generalization ability and prevent overfitting. The complete historical dataset is randomly divided into training and validation sets, with common ratios being 7:3 or 8:2. The training set is specifically used for parameter optimization of the deep neural network model, while the validation set is periodically used during training to evaluate the model's performance on unseen data. Cross-validation can employ the k-fold cross-validation method, dividing the dataset into k subsets, alternating between one subset as the validation set and the remaining k-1 subsets as the training set, repeating training and validation k times, and finally taking the average of the k validation results as an estimate of the model's performance. By observing the changing trends of the loss function on the training and validation sets, it is possible to determine whether the deep neural network model is overfitting. After sufficient training and rigorous cross-validation, the parameters of the deep neural network model are determined, forming a calibrated soil nutrient consumption simulation model that can accurately simulate the dynamic consumption of soil nutrients under specific farmland conditions. Refer to Table 1, which shows a table of weight coefficients and baseline offset values; these values ​​need to be calibrated and determined based on historical data for the specific application scenario.

[0066] Table 1: Parameters for Calculating the Anomaly Index

[0067] Parameter name Symbolic representation Example values illustrate Pearson correlation coefficient weights 0.6 Importance weights of linearly correlated components Spearman rank correlation coefficient weights 0.4 Importance weights of monotonic correlation components baseline offset -0.1 Correction based on correlation with historical normal periods

[0068] It is understandable that the components of the anomaly index need to be adjusted according to the specific application environment. In some embodiments, the weighting coefficients of the Pearson correlation coefficient and the Spearman rank correlation coefficient are not constant. The weighting coefficients can be dynamically optimized by analyzing the relationship between historical fertilization anomalies and the two correlation coefficients, and by using machine learning methods such as logistic regression, so that the anomaly index is more sensitive to real anomalies. The calculation of the baseline offset may need to consider the seasonality of the data, for example, by establishing different baseline offset profiles for different seasons to more accurately reflect the periodic changes in normal background values. The update frequency of the dynamic threshold can be set, for example, by recalculating the anomaly index sequence statistics within a preset period daily or weekly, to ensure that the threshold can track the slow changes in the system state.

[0069] Optionally, the structure of the deep neural network model can take various forms, such as a multilayer perceptron structure containing multiple fully connected hidden layers, each followed by a non-linear activation function like ReLU; or a long short-term memory network structure to better handle the time-series dependence of soil nutrient consumption. When training the deep neural network model, in addition to using the traditional mean squared error as the loss function, a custom loss function designed specifically for agricultural applications can be introduced, such as assigning a higher penalty weight to the prediction error of nutrient deficiency. Early stopping can be used in the training process, terminating training prematurely when the validation set loss no longer decreases to prevent overfitting.

[0070] Optionally, quality control of historical data is crucial. Before using historical soil data, historical environmental observation data, and crop yield data for training, rigorous data cleaning and consistency checks are necessary to remove obvious outliers, handle missing values, and standardize data units and timestamps. If crop yield data originates from yield monitors, its spatial positioning accuracy and data volatility need to be considered, and it may need to be smoothed before being used as training labels. The trained soil nutrient consumption simulation model can be deployed to cloud servers or edge computing devices to achieve real-time inference and prediction of new data.

[0071] See Figure 5The chart displays the calculation results of the fertilization anomaly detection index and dynamic threshold analysis. The chart presents the trajectory of the comprehensive anomaly index in time series form, with the purple curve clearly showing the fluctuations of the anomaly index throughout the monitoring period. The red dashed line represents the detection threshold dynamically calculated by the system according to a preset period. This threshold is determined using the three-sigma criterion based on the mean and standard deviation of the anomaly index sequence, exhibiting good adaptability. The red-filled areas in the chart clearly identify the periods when the anomaly index exceeds the dynamic threshold, and the red scatter dots mark the specific locations of anomaly events, providing agronomists with intuitive decision support. The chart also uses orange and brown dashed lines to show the changing trends of the Pearson correlation coefficient and the Spearman rank correlation coefficient, respectively. These two statistics are core components of the anomaly index. The Pearson correlation coefficient measures the linear correlation between nutrient concentration and precipitation, while the Spearman rank correlation coefficient assesses the monotonic correlation between the two. The weighted fusion of the two ensures the sensitivity of the anomaly detection algorithm to changes in different types of correlations. This multi-indicator fusion method effectively improves the system's accuracy in identifying genuine abnormal events, reduces the risk of false alarms, demonstrates the robustness and practicality of the intelligent fertilization system in complex farmland environments, and provides reliable technical support for sustainable agricultural management.

[0072] Example 5: The process of triggering a multi-level alarm protocol and adjusting fertilizer application based on the identification results of abnormal fertilization events is the core guarantee for the safe operation of the system. The multi-level alarm protocol clearly divides the alarm status into three levels: alert, warning, and severe. Each alarm level corresponds to a clear triggering condition, system response action, and notification strategy. When the application rate of mineral fertilizer per unit area calculated by the fertilizer recommendation algorithm exceeds the preset safe range, the system will initiate the alarm level assessment process. The safe range is a numerical range preset based on crop nutrient requirements, soil environmental capacity, and relevant agricultural environmental standards, with clear upper and lower limits. When the application rate per unit area exceeds the safe range but is within the preset buffer zone, which is a narrow interval adjacent to the boundary of the safe range to avoid frequent alarms caused by small data fluctuations, the system triggers an alert-level alarm. The response action of the alert-level alarm is limited to recording the specific information of this over-limit event in the system operation log, including the time, the application rate per unit area, and the boundary value of the safe range, and sending a prompt message to the responsible agronomist through the system operation interface to remind the agronomist to pay attention to this phenomenon, but the system will not automatically adjust the current fertilizer application rate.

[0073] In practice, when the system detects that the application rate per unit area consistently exceeds the safe range and identifies an abnormal fertilization event through trend consistency analysis, a warning-level alarm is triggered. The determination of continuous exceedance is typically based on a time condition, such as the calculated application rate per unit area exceeding the safe range for two or more consecutive fertilization cycles. Simultaneously, the trend consistency analysis module must confirm the existence of a situation marked as an abnormal fertilization event within the current or recent time window. Once the warning-level alarm conditions are met, the system automatically adjusts the fertilizer application rate, lowering the planned application rate to the upper limit of the safe range. This adjustment is immediate and automatic, requiring no manual intervention. For example, if the upper limit of the safe range is 150 kg per hectare, and the calculated application rate is 170 kg per hectare, the system will limit the actual application rate to 150 kg per hectare. In addition to the adjustment, the system also sends a warning notification SMS to a pre-set agronomist's mobile phone number via an integrated SMS gateway or mobile application push service. The SMS content includes the alarm level, the exceeding value, the adjusted application rate, and the approximate area where the abnormal fertilization event occurred.

[0074] In practice, a severe alarm is triggered when the application rate per unit area significantly exceeds the safe range and is accompanied by abnormal data from key sensors. "Severely exceeding" is typically defined as the application rate per unit area far exceeding the upper limit of the safe range, reaching a threshold that could pose an immediate risk to crops or the environment, such as exceeding the upper limit by 50%. Abnormal data from key sensors refers to abrupt, agronomical changes in sensor readings monitoring soil conductivity, pH, or specific ion concentrations, which may indicate fertilizer blockage, pipe rupture, or concentrated discharge due to misoperation. Upon triggering a severe alarm, the system immediately issues a command to the fertigation unit to stop fertilizer application, completely halting the current fertilization operation. Following this, the system activates a backup irrigation system, independent of the fertilizer mixing pipeline, directly using clean water for irrigation, flushing the farmland to dilute any potentially excessive fertilizer application and mitigate potential harm. While performing the above emergency operations, the system will generate a highest priority notification and send an emergency notification to the agronomist and system administrator. The emergency notification will be sent via both SMS and email, clearly marked as "serious alarm" and requiring immediate manual intervention and on-site verification.

[0075] It is understandable that the effective execution of multi-level alarm protocols relies on precise parameter settings and reliable data sources. In some embodiments, the upper and lower limits of the safety range and the size of the preset buffer are not fixed but dynamically adjusted according to the crop's sensitivity to nutrients at different growth stages. For example, during the seedling stage, the safety range is set more strictly, and the preset buffer is narrower to provide earlier warnings. The criteria for judging abnormal critical sensor data can be based on the rate of change of sensor readings. If a sensor reading changes drastically within a short period, even if the absolute value does not reach an extreme level, it may be considered abnormal and included in the determination of a severe alarm. The notification information template can be customized, including specific field numbers, irrigation zone information, and suggested initial response measures, so that agronomists can quickly understand the situation.

[0076] It is understandable that a smooth transition between alarm levels and the avoidance of oscillations are key design considerations. In some embodiments, the system sets conditions for determining the duration of the alarm state. For example, a warning-level alarm needs to last for a certain period of time (e.g., 24 hours) and the application rate per unit area still has not returned to a safe range before it is upgraded to a warning-level alarm. This can prevent frequent jumps in alarm levels due to temporary data fluctuations. The recovery from a warning-level or severe-level alarm state to a normal state also requires a confirmation process. For example, after the system automatically adjusts the application rate or performs a water rinse, the alarm state will only be lifted after the key sensor data returns to the normal range and the agronomist manually confirms it in the system, allowing the system to resume normal fertilization operations.

[0077] Optionally, alarm log recording and analysis functions are crucial for system optimization. All alarm events, including alert, warning, and critical alarms, are recorded in detail in the system log database. Recorded information includes alarm trigger time, trigger conditions, system actions, notification sending status, and final processing result. This historical alarm data can be used for subsequent analysis, such as identifying recurring alarm patterns in specific fields or assessing the rationality of safety range settings, providing data support for optimizing system parameters.

[0078] Optionally, notification channels can be configured differently based on alarm level and urgency. For alert-level alarms, notifications can be sent only via internal system messages or email; for warning-level alarms, SMS notifications can be added; and for critical-level alarms, a combination of methods such as SMS, automated voice calls, or even strong mobile application alerts can be used to ensure that critical information is received in a timely manner. Multiple notification recipients and their priorities can be set; for example, a critical-level alarm can simultaneously notify frontline agronomists and farm managers.

[0079] Optionally, after triggering a critical alarm and executing the stop fertilizer application and water rinsing operations, the system can automatically generate a brief event report summarizing the time and cause of the event, the measures taken by the system, and recommended follow-up inspection steps. This report can be automatically attached to a notification message or saved in a designated location for agronomists and managers to conduct post-event analysis and archiving. In extreme cases, the system can even be configured to automatically send alarm information to the superior or emergency response team when a critical alarm is triggered and the administrator cannot be contacted, establishing a multi-layered security network.

[0080] 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 method for intelligent dispensing and control of mineral fertilizers for fertigation, characterized in that, The method performs the following steps: Real-time monitoring data on soil moisture, pH, and nutrient concentration are collected through a wireless sensor network set up in farmland; at the same time, environmental observation data on precipitation, temperature, and light intensity are obtained from a remote meteorological server. A soil nutrient consumption simulation model was constructed using the environmental observation data and crop growth calendar, and the parameters of the soil nutrient consumption simulation model were corrected using the real-time monitoring data. Based on the corrected soil nutrient consumption simulation model, a nutrient deficiency distribution map of farmland area was generated using geostatistical methods. The nutrient deficiency distribution map is input into the fertilizer recommendation algorithm to calculate the application rate of mineral fertilizer per unit area, and a trend consistency analysis is performed on the real-time monitoring data and the environmental observation data to identify abnormal fertilization events. When the application rate per unit area exceeds the safe range, a multi-level alarm protocol is triggered and the fertilizer application rate is adjusted based on the identification results of the abnormal fertilization event. Based on the corrected soil nutrient consumption simulation model, the generation of a nutrient deficiency distribution map of farmland area using geostatistical methods includes: extracting spatial coordinates and nutrient values ​​from the real-time monitoring data, calculating a semi-variogram to describe spatial correlation; estimating nutrient levels at unsampled locations using the Kriging interpolation algorithm to generate an initial nutrient distribution surface; comparing the initial nutrient distribution surface with the output of the soil nutrient consumption simulation model, and generating a nutrient deficiency distribution map through iterative optimization. The step of generating a nutrient deficiency distribution map through iterative optimization includes: setting a nutrient deficiency tolerance range; comparing the difference comparison results with the tolerance range, initiating a local densification sampling command for grid areas exceeding the tolerance range, driving a UAV carrying a mobile sensor to perform supplementary measurements; updating the initial nutrient distribution surface using the supplementary measurement data, and repeating the difference comparison step until the differences in all grid areas fall within the tolerance range, and finally outputting the optimized nutrient deficiency distribution map; The nutrient deficiency distribution map is input into a fertilizer recommendation algorithm to calculate the application rate of mineral fertilizer per unit area. Furthermore, a trend consistency analysis is performed between the real-time monitoring data and the environmental observation data to identify abnormal fertilization events, including: Nutrient thresholds are set based on crop type and soil characteristics, and the amount of fertilizer applied at the minimum cost is calculated using linear programming. A sliding window correlation calculation is performed on the nutrient concentration sequence in the real-time monitoring data and the precipitation sequence in the environmental observation data to generate an anomaly index; when the anomaly index exceeds a dynamic threshold, it is marked as a fertilization anomaly event. The generation of the anomaly index includes: calculating the Pearson correlation coefficient and Spearman rank correlation coefficient of the nutrient concentration sequence and the precipitation sequence within a sliding window; weighting and fusing the Pearson correlation coefficient and Spearman rank correlation coefficient, and introducing a baseline offset based on historical normal data to jointly construct the anomaly index; the dynamic threshold is dynamically determined based on the anomaly index sequence of a preset period by calculating its mean and standard deviation, and using the three sigma criterion; the calculation formula for the anomaly index is: in, Indicates an abnormality index. This represents the Pearson correlation coefficient. Represents the Spearman rank correlation coefficient. The weights of the Pearson correlation coefficient are indicated. The weights represent the Spearman rank correlation coefficients. Indicates baseline offset; weight and It is a positive real number and satisfies baseline offset It is a correction term calculated based on the mean of the correlation coefficient sequence during historical normal periods.

2. The intelligent distribution and control method for mineral fertilizers in fertigation as described in claim 1, characterized in that, Real-time monitoring data on soil moisture, pH, and nutrient concentration is collected through a wireless sensor network set up in farmland. This includes deploying multiple sensor nodes in each irrigation zone of the farmland, with each sensor node measuring soil parameters at preset time intervals, and then aggregating the measurement data into a time-series dataset via a gateway device as real-time monitoring data.

3. The intelligent dispensing and control method for mineral fertilizers in fertigation as described in claim 2, characterized in that, The construction of a soil nutrient consumption simulation model using the environmental observation data and crop growth calendar includes: determining the nutrient absorption pattern at different growth stages based on the crop growth calendar, and calculating the natural loss rate of soil nutrients by combining precipitation and temperature data; integrating the nutrient absorption pattern with the natural loss rate to establish a nutrient balance equation based on time steps; and training the nutrient balance equation using historical soil data to form a soil nutrient consumption simulation model.

4. The intelligent dispensing and control method for mineral fertilizers in fertigation as described in claim 3, characterized in that, The calculation of the semivariogram to describe spatial correlation includes: fitting the semivariogram with an exponential model to calculate the nugget value, sill value, and range parameter; selecting the optimal semivariogram model using cross-validation and verifying spatial correlation through spatial autocorrelation analysis; the kriging interpolation algorithm uses ordinary kriging and combines the semivariogram parameters to generate a spatial interpolation weight matrix to estimate the nutrient level at unsampled locations.

5. The intelligent dispensing and control method for mineral fertilizers in fertigation as described in claim 4, characterized in that, The step of training the nutrient balance equation using historical soil data includes: collecting historical soil nutrient data, environmental observation data, and crop yield data for at least one growing season; constructing a deep neural network model, using the nutrient balance equation as the initial model, inputting historical data into the network for supervised learning, and optimizing the model parameters through a backpropagation algorithm; and using a validation set to cross-validate the trained model, thereby forming a calibrated soil nutrient consumption simulation model.

6. The intelligent dispensing and control method for mineral fertilizers in fertigation as described in claim 5, characterized in that, The multi-level alarm protocol for triggering and adjusting fertilizer application based on the identification results of the fertilization anomaly includes: dividing alarm levels into three levels: alert, warning, and severe; when the application rate per unit area exceeds the safe range but is within the preset buffer, an alert-level alarm is triggered, logging is recorded and the agronomist is alerted; when the application rate continues to exceed the safe range and an fertilization anomaly is identified, a warning-level alarm is triggered, the system automatically reduces the fertilizer application rate to the upper limit of the safe range, and sends an SMS notification; when the application rate severely exceeds the safe range and is accompanied by abnormal data from key sensors, a severe-level alarm is triggered, the system immediately stops fertilizer application, starts the backup irrigation system for flushing with clean water, and simultaneously sends an emergency notification to the agronomist and administrator.

Images