Farmland chemical fertilizer application amount analysis and diagnosis method based on machine learning

Abstract

The invention relates to the technical field of farmland fertilization optimization, and discloses a farmland fertilizer application amount analysis and diagnosis method based on machine learning. The method comprises the following steps: firstly, acquiring near infrared spectrum data and surface microscopic image data of farmland soil, and generating a soil comprehensive risk area; analyzing the change gradient of the migration rate of the nitrogen element in the region to evaluate the offset state, and analyzing the interaction strength of the phosphorus element and the potassium element to evaluate the asymmetric coupling state of the phosphorus element and the potassium element; calculating the farmland fertilizer demand degree according to the evaluation result; analyzing the matching deviation between the operation behavior of the peasant household and the fertilizer demand degree in combination with historical fertilization operation records, and generating an operation deviation value; a fertilization optimization strategy is generated in combination with the fertilizer demand degree and the operation deviation value, and execution parameters of the fertilizer applicator are dynamically adjusted through a self-adaptive control algorithm. According to the method, the multi-dimensional information of the soil can be synthesized, the fertilizer demand condition is accurately evaluated, the fertilization strategy is optimized, and the fertilization operation is dynamically adapted.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of farmland fertilization optimization, in particular to a farmland fertilizer application amount analysis and diagnosis method based on machine learning. BACKGROUND

[0002] In current farmland production, the fertilizer application link has long been plagued by the problem of insufficient precision. Traditional fertilization modes often take farmers' past experience as the core, or refer to regional fertilization manuals for operation. This approach is difficult to meet the differentiated needs of different soil characteristics. Soil, as the basic carrier for crop growth, is not uniformly stable in the distribution of key nutrients such as nitrogen, phosphorus, and potassium within it, but is dynamically changing with factors such as tillage cycle, irrigation method, and climate change. For example, after the rainy season, nitrogen may migrate to the deep soil due to leaching, causing a sharp drop in the nutrient content of the surface soil; and in the dry season, the availability of phosphorus is reduced due to the enhanced adsorption of soil colloids. These complex change patterns are difficult to accurately grasp through experiential judgment.

[0003] Existing soil detection technologies, although to some extent, help with fertilization decision-making, still have obvious shortcomings. Near-infrared spectroscopy technology is widely used in soil composition analysis due to its rapid detection advantage, but most of its application scenarios only focus on the determination of the content of a single element, such as the determination of soil total nitrogen or available phosphorus content, ignoring the interaction between different elements. In fact, the absorption efficiency of phosphorus is significantly affected by the concentration of potassium ions, and there is a competitive relationship between the two on the adsorption sites of soil colloids. This interaction is directly related to the actual utilization effect of fertilizers, and existing technologies have limited research on such correlations. At the same time, the microstructure of the surface soil, such as porosity and aggregate distribution, affects the diffusion rate of fertilizers in the soil and the absorption path of crop roots, but the current detection system rarely includes surface micro features in the calculation basis for fertilization amount.

[0004] In actual fertilization operations, farmers lack accurate knowledge of the real-time nutrient status of the soil, which can lead to a mismatch between the amount of fertilizer applied and the actual demand. Some farmers tend to overuse fertilizers in pursuit of high yields, which not only leads to low fertilizer utilization and increased production costs, but also causes nutrients that are not absorbed to seep into groundwater or enter the atmosphere, leading to a series of ecological and environmental problems such as soil acidification and water eutrophication. Some farmers may also experience insufficient fertilization due to concerns about costs or lack of scientific guidance, leading to a shortage of nutrients needed for crop growth, which in turn affects the normal development and yield quality of crops. In addition, once a traditional fertilization plan is established, it is difficult to adjust flexibly according to the dynamic changes of soil nutrients during the crop growth cycle, and it cannot meet the development needs of modern agriculture for precision and intelligent management. SUMMARY

[0005] The present application aims to provide a machine learning-based farmland fertilizer application amount analysis and diagnosis method to solve the problems raised in the background art.

[0006] To achieve the above-mentioned purpose, the present application provides a machine learning-based farmland fertilizer application amount analysis and diagnosis method, which comprises:

[0007] Collecting near-infrared spectrum data and surface microscopic image data of farmland soil to generate a soil comprehensive risk area;

[0008] Analyzing the change gradient of nitrogen element migration rate in the soil comprehensive risk area to evaluate the deviation state of nitrogen migration rate;

[0009] Analyzing the interaction strength of phosphorus and potassium elements in the soil comprehensive risk area to evaluate the asymmetric coupling state of phosphorus and potassium interaction;

[0010] According to the evaluation results of the deviation state of nitrogen migration rate and the asymmetric coupling state of phosphorus and potassium interaction, the fertilizer requirement degree of farmland is calculated;

[0011] Based on historical fertilization operation records, the matching deviation of farmer's operation behavior and the fertilizer requirement degree of farmland is analyzed to generate a farmer's operation deviation value;

[0012] Combining the fertilizer requirement degree of farmland and the farmer's operation deviation value, a fertilization optimization strategy is generated;

[0013] According to the fertilization optimization strategy, the self-adaptive control algorithm is used to dynamically adjust the execution parameters of the fertilizer machine.

[0014] Preferably, the soil comprehensive risk area is generated by:

[0015] Using a random forest regression algorithm to process the near-infrared spectrum data to generate soil nutrient content distribution information;

[0016] Extracting soil particle structure features in the surface microscopic image data through a convolutional neural network model to generate a soil structure feature map;

[0017] Based on the soil nutrient content distribution information and the soil structure feature map, a soil comprehensive risk area is generated through spatial superposition analysis.

[0018] Preferably, the random forest regression algorithm is used to process the near-infrared spectrum data to generate soil nutrient content distribution information, which specifically comprises:

[0019] Pretreating the near-infrared spectrum data to extract effective band reflectivity data;

[0020] The effective band reflectivity data are trained and modeled by using a random forest regression algorithm to predict the content values of organic matter, nitrogen, phosphorus and potassium elements in the soil.

[0021] The spatial distribution of soil nutrient content distribution information is generated according to the predicted soil element content values.

[0022] Preferably, the soil particle structure features in the surface microscopic image data are extracted by the convolutional neural network model to generate a soil structure feature map, specifically including:

[0023] The surface microscopic image data are subjected to noise reduction and contrast enhancement processing;

[0024] The porosity, aggregate morphology and arrangement direction features of the soil particles are extracted by using a pre-trained convolutional neural network model;

[0025] The soil structure feature map covering the farmland area is generated according to the porosity, aggregate morphology and arrangement direction features.

[0026] Preferably, the soil comprehensive risk area is generated by spatial superposition analysis based on the soil nutrient content distribution information and the soil structure feature map, specifically including:

[0027] The spatial coordinates of the element content abnormal area in the soil nutrient content distribution information are identified;

[0028] The spatial coordinates of the structure defect area in the soil structure feature map are identified;

[0029] The element content abnormal area and the structure defect area are subjected to spatial overlay operation to extract the boundary of the overlapping area to generate the soil comprehensive risk area.

[0030] Preferably, the change gradient of the nitrogen element migration rate in the soil comprehensive risk area is analyzed to evaluate the deviation state of the nitrogen migration rate, specifically including:

[0031] The nitrogen element concentration monitoring data at multiple time points in the soil comprehensive risk area are obtained;

[0032] The gradient change amount of the nitrogen element concentration at adjacent time points is calculated to generate a nitrogen migration rate gradient sequence;

[0033] By comparing the nitrogen migration rate gradient sequence with a preset threshold range, it is determined whether the deviation state of the nitrogen migration rate is normal.

[0034] Preferably, the interaction strength of the phosphorus element and the potassium element in the soil comprehensive risk area is analyzed to evaluate the asymmetric coupling state of the phosphorus-potassium interaction, specifically including:

[0035] Synchronously collect spectral response data of phosphorus and potassium elements in the soil comprehensive risk area;

[0036] Calculate the correlation strength of spectral response of phosphorus and potassium elements by using mutual information algorithm;

[0037] According to the direction difference of the correlation strength, determine whether the asymmetric coupling state of phosphorus-potassium interaction is reasonable.

[0038] Preferably, the fertilizer requirement degree of the farmland is calculated according to the evaluation results of the deviation state of the nitrogen migration rate and the asymmetric coupling state of the phosphorus-potassium interaction, and specifically includes:

[0039] When the deviation state of the nitrogen migration rate is abnormal, mark a first risk weight coefficient;

[0040] When the asymmetric coupling state of the phosphorus-potassium interaction is unreasonable, mark a second risk weight coefficient;

[0041] The first risk weight coefficient and the second risk weight coefficient are weighted and fused to output a quantitative value of the fertilizer requirement degree of the farmland.

[0042] Preferably, the matching deviation between the farmer's operation behavior and the fertilizer requirement degree of the farmland is analyzed based on the historical fertilization operation records to generate a farmer's operation deviation value, and specifically includes:

[0043] Extract time, location and fertilization amount data in the historical fertilization operation records;

[0044] Compare the quantitative value of the fertilizer requirement degree of the farmland with the actual fertilization amount data at the same location;

[0045] Calculate the standard deviation ratio of the fertilizer requirement amount and the actual fertilization amount to generate the farmer's operation deviation value.

[0046] Preferably, the fertilizer optimization strategy is generated by combining the fertilizer requirement degree of the farmland and the farmer's operation deviation value, and specifically includes:

[0047] Generate a basic fertilization amount matrix according to the quantitative value of the fertilizer requirement degree of the farmland;

[0048] Modify the basic fertilization amount matrix by introducing the farmer's operation deviation value;

[0049] Output the fertilizer optimization strategy including fertilization location, fertilization timing and amount;

[0050] The adaptive control algorithm is used to dynamically adjust the execution parameters of the fertilizer machine according to the fertilizer optimization strategy, and specifically includes:

[0051] Real-time acquisition of the traveling position of the fertilizer applicator and soil moisture monitoring data, matching the target fertilizer amount of the current position based on the fertilizer optimization strategy, and dynamically adjusting the discharging rate and spraying pressure parameters of the fertilizer applicator by using a fuzzy PID controller.

[0052] Compared with the prior art, the beneficial effects of the present application are:

[0053] By collecting the near-infrared spectrum data and surface microscopic image data of farmland soil, the comprehensive characteristics of the soil can be captured from the macro and micro levels, providing a comprehensive basis for subsequent risk area division. Combined with the nutrient content characteristics contained in the spectrum information and the soil structure information reflected by the surface microscopic image, the generated soil comprehensive risk area is more in line with the actual farmland conditions, which helps to accurately lock the areas that need to be paid attention to.

[0054] When analyzing the change gradient of the nitrogen element migration rate, the limitations of traditional static detection are broken through, and the movement law of nitrogen elements in the soil is dynamically tracked, so as to accurately evaluate the deviation state. This dynamic analysis method can timely find the abnormal distribution of nitrogen elements, and provide direction for reasonable regulation of nitrogen fertilizer application amount.

[0055] For the analysis of the interaction strength of phosphorus and potassium elements, the synergistic or antagonistic relationship between the two elements in the soil is focused on, and the asymmetric coupling state is evaluated, avoiding the judgment deviation that may occur when considering a single element alone. Understanding the interaction mechanism of phosphorus and potassium elements can more scientifically formulate the matching scheme of the two elements and improve the overall utilization efficiency of fertilizers.

[0056] According to the nitrogen migration rate deviation state and the phosphorus-potassium asymmetric coupling state, the farmland fertilizer requirement degree is calculated, realizing the quantification and precision of fertilizer requirement evaluation, and getting rid of the extensive mode relying on experience estimation in the past. Based on the historical fertilization operation record, the matching deviation of the farmer's operation behavior and the farmland fertilizer requirement degree is analyzed, and the generated farmer operation deviation value can reflect the problems existing in the actual fertilization process, providing a reference for correcting unreasonable operation.

[0057] The fertilization optimization strategy generated by combining the farmland fertilizer requirement degree and the farmer operation deviation value takes into account the actual demand of the soil and the operation habit of the farmer, and has stronger practicability and operability. Through the adaptive control algorithm, the fertilizer applicator execution parameters can be dynamically adjusted, so that the fertilization operation can respond to the change of soil nutrients in real time, ensuring that the fertilizer application amount is always within a reasonable range, which meets the needs of crop growth, reduces unnecessary fertilizer consumption, and reduces the negative impact on the ecological environment. BRIEF DESCRIPTION OF DRAWINGS

[0058] Figure 1 The timing diagram of the farmland fertilizer application amount analysis and diagnosis method based on machine learning described in the present application;

[0059] Figure 2 Flowchart for generating soil comprehensive risk area;

[0060] Figure 3 Flowchart for generating risk area for spatial overlay analysis;

[0061] Figure 4 Flowchart for evaluating nitrogen migration rate deviation state;

[0062] Figure 5 Flowchart for generating farmer operation deviation value. DETAILED DESCRIPTION

[0063] The technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor fall within the scope of protection of the present application.

[0064] Please refer to Figure 1 The present application provides a kind of based on machine learning's farmland fertilizer application amount analysis and diagnosis method, the method comprises:

[0065] Collect near-infrared spectrum data and surface microscopic image data of farmland soil, generate soil comprehensive risk area using machine learning algorithm. By analyzing the change gradient of nitrogen element migration rate in risk area, its deviation state is evaluated;At the same time, the interaction intensity of phosphorus and potassium elements is analyzed, and the asymmetric coupling state is quantified. Combine the above evaluation results to calculate the fertilizer requirement degree of farmland, and generate the farmer operation deviation value by comparing with the historical fertilization operation record. Finally, the fertilizer optimization strategy is generated by comprehensively generating the fertilizer requirement degree and the operation deviation value, and the variable precision fertilization is realized by dynamically adjusting the execution parameters of fertilizing machine through adaptive control algorithm.

[0066] Example 1: refer to Figure 2 The near-infrared spectrum data is collected by using a high-precision portable spectrometer, and the wavelength range is 900-2500nm, and the spectral resolution is 3nm. During the collection process, the probe of the spectrometer and the surface of the soil keep a vertical distance of 10cm, and the spectral data of each sampling point is collected 3 times and the average value is taken to reduce random error. The original spectrum data after collection is first corrected for dark current to eliminate environmental light interference, and then calibrated for reflectivity using a standard white plate. In the preprocessing stage, Savitzky-Golay filter is used to smooth noise, with window size of 11 data points and polynomial order of 2 to retain effective spectral characteristics while suppressing high-frequency noise. Standard normal variable transformation is used to eliminate soil particle scattering effect, so that the spectral data of different sampling points are comparable.

[0067] The preprocessed effective band reflectance data is input into a random forest regression model for training and prediction. The model consists of 500 decision trees, with a maximum depth of 15 layers for each tree to prevent overfitting. The input features are the reflectance values in the 1350-2500 nm range, which is sensitive to the absorption characteristics of soil organic matter, nitrogen, phosphorus, potassium, and other nutrients. The model training uses ten-fold cross-validation, with 90% of the data randomly selected for training and the remaining 10% for validation to evaluate the model's generalization ability. After training, the model outputs the predicted values of soil organic matter content, total nitrogen content, available phosphorus content, and available potassium content. The prediction results are spatially interpolated using the Kriging interpolation method to generate a 1 m x 1 m resolution soil nutrient content distribution raster map, which contains the nutrient content information of each pixel point for subsequent generation of soil comprehensive risk areas.

[0068] The ground micro-image data is collected by a high-resolution micro-camera with a resolution of 5 million pixels, equipped with a ring LED light source to reduce shadow interference. During collection, the camera maintains a fixed distance of 5 cm from the soil surface, and 3 images are taken at each sampling point with different angles of illumination to enhance texture contrast. In the image preprocessing stage, bilateral filtering is used to remove noise, with a spatial domain standard deviation of 3 pixels and a gray value domain standard deviation of 10 to preserve edge information while denoising. Limited contrast adaptive histogram equalization is used to enhance image contrast, with a block size of 16 x 16 pixels and a contrast limit coefficient of 2.0 to prevent local over-enhancement.

[0069] The preprocessed micro-image is input into a pre-trained ResNet50 convolutional neural network to extract soil particle structure features. The network is pre-trained on the ImageNet dataset and fine-tuned on the soil micro-image dataset to adapt to the feature extraction task of soil particles. The network output includes porosity, aggregate morphology, and arrangement direction features. Porosity is obtained by calculating the proportion of pore area in the entire image. Aggregate morphology features include aspect ratio, circularity, and area parameters, which are calculated after segmenting the aggregate outline using the U-Net network. The arrangement direction feature is quantified using the histogram of oriented gradients method, which counts the distribution of soil particles in different directions.

[0070] The extracted soil structure features are geocoded to generate a soil structure feature map. The map uses the same spatial reference system as the soil nutrient content distribution map to ensure accurate spatial matching. Each pixel point in the map records the porosity, aggregate morphology parameters, and arrangement direction features, which are used for subsequent identification of soil comprehensive risk areas.

[0071] The spatial overlay analysis of soil nutrient content distribution information and soil structure feature map was completed by geographic information system software. The nutrient content abnormal area was identified by Z-Score standardization method, wherein the area with nitrogen content lower than 1.5 times of the standard deviation of the mean value or the area with phosphorus and potassium content higher than 2 times of the standard deviation of the mean value was marked as abnormal. The soil structure defect area was extracted by Otsu threshold segmentation method, wherein the area with porosity lower than 15% or the area with aggregate circularity higher than 0.85 was marked as structure defect area. After the spatial coordinates of the two types of abnormal areas were converted into UTM projection, the intersection operation in overlay analysis was used to extract the overlapping area. The morphological closing operation was performed on the overlapping area to eliminate small cavities, and the structure element size was 3*3 pixels. Finally, the Douglas-Peucker algorithm was used to simplify the boundary polyline, the tolerance was set to 0.1 m, and the soil comprehensive risk area in vector format was generated. The attribute table of the area recorded the nutrient abnormality level and structure defect type of each polygonal unit, which provided the basis for subsequent element migration analysis and fertilization decision.

[0072] Example 2: see Figure 3 After the generation of the soil nutrient content distribution information, standardization processing was required to identify the abnormal area. The distribution information was stored in the form of a grid, and each pixel point contained the predicted content values of organic matter, total nitrogen, available phosphorus and available potassium. Z-Score standardization method was used to normalize each nutrient index, and the standardized value of each pixel point was calculated. For nitrogen element, the area with standardized value lower than 1.5 times of the standard deviation of the mean value was marked as nitrogen content abnormal area; for phosphorus and potassium elements, the area with standardized value higher than 2 times of the standard deviation of the mean value was marked as phosphorus and potassium content abnormal area. The boundary of the abnormal area was extracted by edge detection algorithm, and Canny operator was used for edge recognition, wherein the kernel size of Gaussian filter was 5*5, and the ratio of high threshold value to low threshold value was set to 3:1.

[0073] The analysis of soil structure feature map focused on two key indicators: porosity and aggregate morphology. The porosity feature was calculated by image segmentation technology, and Otsu adaptive threshold method was used to determine the segmentation threshold of pores and soil particles. For the area with porosity lower than 15%, it was determined as structure compacted area; for the area with aggregate circularity higher than 0.85, it was determined as aggregate morphology abnormal area. The formula for calculating circularity is as follows:

[0074]

[0075] wherein: represents the circularity of the aggregate, represents the area of the aggregate, represents the perimeter of the aggregate. The closer the circularity is to 1, the closer the aggregate morphology is to the ideal circular shape. The spatial distribution of structure defect area was determined by connected component analysis, and the minimum connected component area was set to 0.01 m² to filter noise interference.

[0076] The spatial coordinates of the nutrient anomaly area and the structural defect area are uniformly converted to the UTM projection coordinate system to ensure consistent spatial reference. The overlay analysis is completed using the spatial operation module of the geographic information system software. The specific steps include: first, rasterizing the two types of abnormal areas with a resolution of 0.5m x 0.5m; then performing a logical "and" operation to extract the overlapping part of the two types of abnormal areas; finally, converting the overlapping area to a vector polygon. During the overlay process, the bilinear interpolation method is used to handle the resampling of raster data to reduce geometric distortion.

[0077] The initial overlapping area generated may have small holes or jagged boundaries, which need to be post-processed morphologically. The closing operation is used to fill small holes in the area, and the structural element is selected as a circular kernel with a diameter of 3 pixels. The boundary simplification uses the Douglas-Peucker algorithm with an iteration tolerance of 0.2m, which retains the main geometric features while reducing the number of vertices. The topology of the simplified polygon needs to be checked to ensure that there are no self-intersection or hanging node topological errors.

[0078] The final generated soil comprehensive risk area is stored in vector polygon format, and the attribute table contains the following fields: polygon unique identifier, area, average nitrogen content, average phosphorus content, average potassium content, porosity level, and aggregate form type. Among them, the porosity level is divided into low (<15%), medium (15%-30%), and high (>30%); the aggregate form type is divided into regular type (circularity ≤0.85) and irregular type (circularity >0.85). Spatial data is output in GeoJSON format for easy integration with subsequent analysis modules.

[0079] In the data visualization section, the soil comprehensive risk area is rendered using the hierarchical color setting method. The nutrient anomaly area is represented by the red color system, and the color depth reflects the degree of abnormality; the structural defect area is represented by the blue color system, and the color tone lightness corresponds to the porosity level. The overlapping part of the two areas is identified by purple, highlighting the high-risk area. The visualization result supports zooming and panning operations, with a minimum display scale of 1:5000 to ensure the accuracy requirements of field management operations.

[0080] The spatiotemporal change analysis of the risk area is achieved through time series comparison. For multiple monitoring data of the same land, the change detection technology is used to identify the expansion or contraction trend of the risk area. Change detection is based on the raster difference method, and the difference threshold is set to a change in nutrient content of more than 10% or a change in porosity of more than 5%. The detection results are displayed in animation form, and the time interval can be adjusted to facilitate observation of short-term dynamic changes.

[0081] The generation process of the comprehensive risk area involves a large amount of spatial data processing, and the optimization of computational efficiency is the key. The grid operation adopts a block processing strategy, with a block size of 1024x1024 pixels. The excess part is automatically divided into multiple tasks for parallel computing. Spatial index is applied to accelerate the query of vector data processing, and the node capacity of R-tree index is set to 50. Distributed computing framework is used to process large-scale farmland data, and the master-slave architecture is adopted for task scheduling, with the master node responsible for task allocation and result aggregation, and the slave node executing specific operations.

[0082] The attribute data analysis of the risk area uses spatial statistical methods. Global spatial autocorrelation is evaluated by Moran's I index to assess the spatial clustering of nutrient content and structural characteristics. Local hotspot analysis uses Getis-Ord Gi* statistics to identify significant high-value or low-value clusters. The statistical results are overlaid with the risk area boundaries to verify the reasonableness of the area division. The statistical significance level is set to 0.05, and the Bonferroni correction is used to control the multiple comparison error.

[0083] The output data interface design of the comprehensive risk area of soil considers the compatibility of multiple systems. Spatial data services follow OGC standards and support WMS and WFS protocols; attribute data is provided in the form of RESTful API, returning JSON format query results. The data caching mechanism uses the LRU algorithm, with a cache size of 1GB, and the most recently used data is prioritized for retention. User permission management is based on role-based access control, with three levels of administrator, agronomist, and operator, ensuring data security and operation traceability.

[0084] Special attention is needed for the processing of farmland boundary data. Before overlay analysis, the soil sampling area and the actual boundary of the farmland need to be accurately matched. The boundary matching uses the minimum bounding rectangle method, and the areas with a deviation of more than 0.5m are geometrically corrected. The correction method is based on control point transformation, and control points are placed at the four corners of the farmland. The affine transformation model is used to adjust the coordinate system. The corrected boundary is cropped with the risk area data to remove invalid data outside the farmland, ensuring the accuracy of the analysis results.

[0085] The update mechanism of the risk area is designed for incremental processing. When new soil sampling data is added, only the changed area is recalculated for nutrient content and structural characteristics, avoiding the resource consumption of full update. The change area detection is based on spatio-temporal index, and the index structure uses a hybrid index of quadtree spatial division and B+ tree temporal division. The triggering conditions for incremental update tasks include: the coverage of new sampling data exceeds 10%, or the distance from the last update is more than 30 days. The update process records operation logs, including timestamp, processing range, and calculation parameters, supporting backtracking analysis.

[0086] The verification of the comprehensive risk zones of soil was performed by an independent sampling point evaluation method. Verification sampling points were deployed inside and outside the generated risk zones, with a sampling density of one point per 10 mu. The verification data included laboratory-measured nutrient content and micro-image-analyzed structural parameters. The evaluation indicators were calculated using a confusion matrix to evaluate producer accuracy and user accuracy, where producer accuracy reflects the identification ability of the risk zones for actual problems, and user accuracy represents the reliability of the zone division. The verification results were used to dynamically adjust the risk determination threshold to form a closed-loop optimization.

[0087] The application scenarios of the risk zones are not limited to fertilization decisions. In irrigation management, the comprehensive risk zones can be used to identify abnormal water infiltration areas; in the development of tillage schemes, structural defect areas suggest areas that need to be deep loosened or improved. Multi-scene applications are achieved through metadata identification, with each risk zone object attached to a scene tag. The system automatically matches the corresponding analysis model and disposal recommendations according to the tag. Scene tags are organized using an ontology method, supporting semantic reasoning and intelligent recommendations.

[0088] The correlation analysis of farmland management historical data and risk zones uses a spatiotemporal matching method. The spatial and temporal superposition of fertilization records, yield data, and weather information over the years with the risk zones is used to mine the influence of long-term management patterns on soil conditions. Spatiotemporal matching is based on GeoHash encoding, which unifies two-dimensional space and one-dimensional time into a string prefix to accelerate spatiotemporal queries. The analysis results are used to optimize the generation parameters of the risk zones, such as adjusting the time decay coefficient for nutrient anomaly determination, making the model more consistent with the actual farmland evolution trend.

[0089] The interactive analysis tool of the comprehensive risk zones of soil supports multi-dimensional data exploration. Users can adjust the determination thresholds of nutrient anomalies and structural defects through sliders and observe the range changes of the risk zones in real time. Parallel coordinate plots are used to display the relationships between multiple indicators, helping to discover potential synergies or antagonistic effects. The event response delay of interactive operations is controlled within 200 ms, and background calculations are implemented through WebWorkers to not block front-end interactions. User operation history is automatically saved, supporting rollback and scheme comparison to assist the decision-making process.

[0090] Embodiment 3: see Figure 4The soil comprehensive risk area is laid out when the time domain reflection sensor network is used for nitrogen element migration monitoring. The sensor nodes are arranged in a 20 m x 20 m grid, and each node monitors the nitrogen concentration of the 0-30 cm soil layer with a sampling interval of 2 hours. The sensor data is transmitted to the base station through the LoRa wireless network, and the transmission power is set to 14 dBm, and the air rate is selected as 292 bps to balance the transmission distance and energy consumption. The original monitoring data is first subjected to outlier rejection, and the Tukey method is used to identify outliers, with a quartile range coefficient of 3.0. Missing data is completed by cubic spline interpolation, with an interpolation node interval of 4 hours to ensure curve smoothness while avoiding overfitting.

[0091] The calculation of nitrogen migration rate is based on the time gradient of concentration change. For the nitrogen concentrations of adjacent time points and , the calculation formula of migration rate v is:

[0092]

[0093] Among them: represents the nitrogen migration rate (mg / (L·h)), and respectively represent the nitrogen concentrations (mg / L) at and time, is the time interval (h). The calculated rate sequence is filtered by moving average, with a window width of 3 consecutive data points to suppress short-term fluctuation interference.

[0094] The evaluation of the migration rate gradient sequence uses a dynamic threshold method. The preset threshold range is set according to the soil texture classification: sandy soil is 0.05-0.2 mg / (L·h), loam is 0.03-0.15 mg / (L·h), and clay is 0.02-0.1 mg / (L·h). Soil texture information is determined by laser diffraction method, and input parameters include the percentage of clay particles (<0.002 mm), silt particles (0.002-0.05 mm), and sand particles (0.05-2 mm). When the migration rate of three consecutive time points exceeds the threshold range, or the adjacent rate change amplitude exceeds 30%, it is determined to be an excursion state anomaly. The anomaly event record includes the occurrence time, duration, and maximum deviation amplitude, which is used for subsequent risk weight calculation.

[0095] ​​The phosphorus-potassium interaction analysis was performed by collecting data synchronously using LIBS technique. The spectral range of the portable LIBS spectrometer was 200-900 nm with a resolution of 0.1 nm. The characteristic spectral lines of potassium KⅠ766.5 nm and phosphorus PⅠ213.6 nm were preprocessed, including dark current subtraction, wavelength calibration, and intensity normalization. The spectral line analysis was performed by fitting Voigt function, considering the comprehensive effect of instrument broadening and Doppler broadening, and the goodness of fit was required to be R²>0.95. The peak area integration was calculated by the trapezoidal method, and the baseline correction was achieved by the iterative least squares method, with the iteration number limited to 10 times.

[0096] The mutual information algorithm was used to quantify the correlation strength of the phosphorus-potassium spectral response. The phosphorus-potassium peak area sequence of the continuous 5 sampling points was taken as the input, and the window width parameter h of the kernel density estimation was set to 0.05. The joint probability distribution was calculated by the Epanechnikov kernel function, and the integral interval was discretized into 100 equal parts. The mutual information value I(P;K) of phosphorus to potassium and the mutual information value I(K;P) of potassium to phosphorus were calculated, and the directionality difference D was defined as:

[0097]

[0098] When the absolute value of D exceeded 0.15, it was determined that there was significant asymmetric coupling between phosphorus and potassium. The coupling state classifier output three results: phosphorus dominant type (D>0.15), potassium dominant type (D<-0.15), or balanced type (|D|≤0.15). The classification results were recorded together with auxiliary parameters such as soil pH value and organic matter content to form an interaction feature vector.

[0099] The spatiotemporal correlation analysis of nitrogen migration bias state and phosphorus-potassium coupling state used the colocalization detection method. For each 1m×1m grid cell, its nitrogen migration anomaly label and phosphorus-potassium coupling type were extracted. The spatial autocorrelation analysis used the JoinCount statistic to calculate the proportion of adjacent cells with the same state. The temporal synchronicity was evaluated by the cross-correlation function, with a maximum time lag of 12 hours. The analysis results were visualized using a heat map matrix, with the horizontal axis representing the nitrogen migration anomaly intensity and the vertical axis representing the phosphorus-potassium coupling directionality difference, and the color scale reflecting the co-occurrence frequency.

[0100] The quality control of monitoring data was carried out throughout the entire implementation process. The sensor was calibrated regularly at two points, with deionized water for low point calibration and 100 mg / L potassium nitrate standard solution for high point calibration. The LIBS spectrometer was calibrated daily using the characteristic spectral lines of a mercury-argon lamp as a reference.

[0101] The data transmission and storage adopt a hierarchical architecture. The edge computing node is responsible for raw data preprocessing and real-time alarm, deployed on the field gateway device. The processed data is uploaded to the cloud platform and stored in the time series database. The data point retention policy is set to retain the original data for 30 days, the hourly average for 1 year, and the daily average permanently. The query interface supports time range filtering and spatial range screening, with a response time requirement of less than 1 second. Data backup adopts the off-site multiple copy strategy, with daily incremental backup and weekly full backup.

[0102] The dynamic updating mechanism of the nitrogen migration model is based on online learning. Model retraining is triggered when the monthly new data reaches 10% of the total. Recursive feature elimination is used for feature selection, retaining the top 80% of important features. Model hyperparameters are adjusted through Bayesian optimization, with 50 iterations. The updated model is verified through A / B testing, with the new and old models running in parallel for a week. The version with lower mean absolute error is selected for online use. Model version management uses semantic version control, with major updates incrementing the major version number.

[0103] The visualization tool for phosphorus-potassium interaction analysis supports multi-dimensional data exploration. The three-dimensional scatter plot has X-axis representing phosphorus content, Y-axis representing potassium content, Z-axis representing directional difference degree, and point color representing soil pH. The parallel coordinate chart shows the relationship between interaction intensity and soil texture, organic matter, etc. Users can focus on specific data subsets through filtering operations, and the system dynamically updates statistical summaries. Visualization rendering uses WebGL acceleration, supporting smooth interaction with 100,000 data points.

[0104] The decision rules for anomaly diagnosis use fuzzy logic systems. Input variables include nitrogen migration rate deviation, phosphorus-potassium coupling asymmetry, and soil moisture index, each divided into 3 linguistic variables. The fuzzy rule base contains 27 IF-THEN rules, such as "IF high nitrogen migration deviation AND strong phosphorus-potassium asymmetry AND low moisture THEN high risk level." Defuzzification uses the gravity method, outputting a risk score of 0-1. The rule base can be maintained through an expert knowledge editing interface, with modification records stored in a version control system.

[0105] The energy management of the monitoring system uses an adaptive strategy. The sampling frequency is dynamically adjusted based on battery remaining capacity and recent data fluctuation. The base frequency is 2 hours / once, which can be increased to 30 minutes / once when detecting abnormal trends, and reduced to 4 hours / once in stable state. Data compression algorithms are used for wireless transmission, with a compression ratio controlled between 1:5 and 1:10. Solar panels power the field nodes, with supercapacitors to cope with rainy weather, ensuring continuous 7-day operation without sunlight.

[0106] The algorithm optimization of spatio-temporal data analysis is designed for the characteristics of farmland. The spatial interpolation adopts the improved Kriging method, introducing soil type as an auxiliary variable. The time series prediction combines seasonal decomposition and LSTM neural network, with a sliding window width of 7 days. The distributed computing framework processes large-scale data, using the MapReduce model, where Mapper is responsible for spatial partitioning calculation and Reducer performs time dimension aggregation. The task scheduling considers data locality, prioritizing tasks to computing nodes that store the required data.

[0107] The user interface is divided into three view layers: the real-time monitoring view displays the current sensor network status and the latest alarm information; the historical analysis view provides multi-period data comparison and trend charts; the decision support view presents risk assessment results and disposal suggestions. The interface responsive design adapts to various devices from mobile phones to desktops, with touch operations supporting pinch zoom and swipe viewing. The permission control system assigns function modules based on roles, allowing agronomists to configure algorithm parameters, technicians to view device status, and farmers to receive simplified alarm notifications.

[0108] System integration testing covers the entire process verification. Simulation testing generates a dataset containing typical abnormal patterns to verify the sensitivity of the detection algorithm. Stress testing simulates 500 nodes uploading data simultaneously to test the system throughput. Field testing selects plots of three typical soil types and runs continuously for one complete crop growing season to record system stability and data quality. Test results form an improvement report to guide algorithm parameter adjustment and hardware selection optimization.

[0109] The maintenance and upgrade mechanism ensures long-term reliable operation of the system. Remote diagnosis tools can detect common problems such as node offline and data anomalies. Firmware upgrades use differential update technology, only transmitting the changed part to reduce bandwidth occupancy. Hardware maintenance plans include cleaning the sensor optical window every quarter, replacing the desiccant and battery every year. Maintenance records are digitally managed, generating device health reports and predictive reminders for replacing aging components.

[0110] Data integration with other agricultural management systems is achieved through standard interfaces. Weather data access uses the AgWeatherNet protocol, and crop growth model input and output comply with the AgML standard. Farm operation data are accessed through an ISOBUS converter, enabling automatic association of fertilizer application records and soil monitoring data. Third-party systems calling the system services need to pass OAuth2.0 authentication, and API access is limited to 100 requests per minute.

[0111] The data analysis report automatic generation module is customized on demand. The daily report contains a summary of abnormal events in the previous 24 hours; the weekly report adds trend analysis and regional comparison; the quarterly report integrates the data characteristics of the entire growth period of the crop. The report template is defined in Markdown, with data automatically filled in, supporting both PDF and HTML output formats. Key indicators are pushed through SMS or mobile applications, and important alarms trigger phone voice notifications to ensure timely handling.

[0112] Example 4: see Figure 5 In the process of assessing the fertilizer requirement degree of farmland, the deviation state of nitrogen migration rate and the asymmetric coupling state of phosphorus and potassium interaction are quantified as risk weight coefficients. When the system detects an abnormal deviation state of nitrogen migration rate, a first risk weight coefficient is assigned according to the magnitude of the deviation from the preset threshold. Specifically, a coefficient of 0.3 is assigned when the deviation magnitude is between 30% and 50%, 0.5 is assigned when it is between 50% and 70%, and 0.7 is assigned when it exceeds 70%. The second risk weight coefficient is generated by the asymmetric coupling state assessment of phosphorus and potassium interaction. When the absolute value of the directional difference D is between 0.15 and 0.25, a coefficient of 0.2 is assigned, between 0.25 and 0.35, a coefficient of 0.3 is assigned, and when it exceeds 0.35, a coefficient of 0.5 is assigned. The two types of coefficients are weighted and fused by the entropy weight method, with the nitrogen state weight accounting for 60% and the phosphorus and potassium interaction weight accounting for 40%. The final output is a quantitative value of the fertilizer requirement degree of farmland in the range of 0-1.

[0113] The historical fertilization operation records are derived from the operation data collected by Internet of Things terminal devices, including GNSS trajectory of fertilization machines, operation timestamps, and real-time discharge amounts. In the data preprocessing stage, signal loss periods and obvious outliers, such as records of single-point fertilization amount exceeding 50 kg / acre, are cleaned up. The system establishes a spatio-temporal index to match fertilization operations with farmland grid cells, with each 5m×5m grid associated with its historical fertilization amount sequence. The comparison analysis of fertilizer requirement and actual fertilization amount uses a sliding window method, with a window size of 3 consecutive operation seasons. The standard deviation ratio of predicted fertilizer requirement to actual fertilization amount in the window is calculated as the farmer's operation deviation value. After exponential smoothing, this deviation value is discretized into 5 correction levels, corresponding to different fertilization amount adjustment magnitudes. See Table 1.

[0114] Table 1: Evaluation results of a typical field.

[0115] Grid number Nitrogen migration offset amplitude (%) Phosphorus-potassium difference degree D First risk weight Second risk weight Fertilizer requirement degree value Average fertilization amount in the past three seasons (kg / mu) Operation deviation level A-12 42.3 0.18 0.3 0.2 0.26 28.5 2 B-07 68.5 0.31 0.7 0.3 0.54 35.2 4 C-15 55.1 -0.22 0.5 0.2 0.38 31.8 3 D-09 72.8 0.41 0.7 0.5 0.62 42.6 5

[0116] The farmland management decision system displays the above evaluation results through a visual interface. The map view uses a layered rendering technique, with the bottom layer showing the farmland boundaries and grid division, the middle layer presenting the fertilizer requirement value in color blocks, with warm colors indicating high fertilizer requirement areas. The upper layer superimposes the historical fertilizer application isopleths, with a line spacing of 5 kg / acre. Clicking on any grid pops up a details panel, showing the risk weight composition, deviation level evolution trend, and recommended adjustment scheme for that location. The interactive filtering tool allows users to filter grids by fertilizer requirement range, deviation level, or soil type, facilitating targeted analysis.

[0117] The deviation correction module dynamically adjusts the fertilization prescription based on the operation deviation level. For grids with levels 1-3, the base fertilization amount is maintained or slightly reduced; for levels 4-5, the fertilization amount is significantly increased, with adjustments to the NPK ratio. The specific adjustment strategy takes into account crop type differences, such as larger nitrogen adjustment than phosphorus and potassium in corn fields, and balanced adjustment of all elements in vegetable fields. The corrected fertilization scheme generates a prescription map, with GeoJSON format storing the fertilizer ratio, amount, and operation time suggestions for each polygon.

[0118] The system implements a complete feedback optimization mechanism. After each fertilization operation is completed, the actual execution data is compared with the recommended value to calculate the degree of agreement. When the regional agreement degree consistently falls below 70%, the retraining process of the deviation evaluation model is triggered. Model updates use incremental learning, preserving the statistical characteristics of historical data but reducing their weight, allowing the system to adapt to changes in farmers' operating habits. Meanwhile, agronomic experts can manually adjust the correction coefficients of specific fields through the management interface, and these manual intervention records are marked and used to verify the effectiveness of the automatic correction strategy.

[0119] Data quality control measures are implemented throughout the entire implementation process. Field sensors are calibrated on-site every month, using standard solutions to verify the accuracy of nitrogen concentration measurements. Fertilizer machine metering devices are verified by the measurement department every quarter to ensure the reliability of the discharge amount records. Data transmission uses encryption and verification mechanisms to prevent tampering or loss during transmission. The database implements row-level security policies, and users with different roles can only view and modify data within their permission scope.

[0120] The system deployment uses a microservices architecture, with functions such as risk assessment, deviation analysis, and prescription generation running as independent services. Containerized deployment ensures resource isolation and elastic expansion of each module, and the Kubernetes cluster automatically manages the number of service instances. Message queues handle high-concurrency data streams, and tasks accumulated during peak periods are executed asynchronously in priority order. The monitoring system tracks the health status of each service, and abnormal conditions trigger automatic restart or failover.

[0121] Mobile application provides a convenient way for farmers to interact. Before operation, the APP pushes the recommended fertilization amount for the day, combined with the current location to automatically display the required fertilizer information of the grid. During operation, the actual fertilization amount is recorded in real time, compared with the recommended value, and the deviation degree is displayed in the form of progress bar. After operation, an execution report is generated, which summarizes the coincidence index and potential improvement suggestions of each field. Offline mode caches the data of recently accessed fields, allowing users to view key information even in areas with poor network signal.

[0122] Compared with traditional agronomic guidance methods, the implementation of this embodiment has the characteristics of dynamic adaptability and spatial specificity. The system not only considers the soil background conditions, but also continuously tracks the deviation between the actual operation of the farmers and the theoretical demand, and continuously optimizes the decision-making through feedback loop. The spatial analysis granularity reaches 5 meters, which can identify small-scale variations within the field and avoid local overuse or underuse caused by overall averaging. This fine management mode is particularly suitable for farmland with large soil variation or long-term uneven fertilization problems.

[0123] The technical challenges in the implementation process mainly come from three aspects: multi-source data spatio-temporal alignment, farmer behavior pattern recognition and system response real-time. Data alignment uses spatio-temporal indexing combined with interpolation method to unify data of different resolutions and sampling frequencies to standard grid. Behavior pattern analysis combines operation sequence mining and machine learning to distinguish systematic deviation and accidental error. Real-time performance is guaranteed through edge computing, which deploys lightweight models on field gateway devices to handle urgent decision-making tasks.

[0124] System maintenance includes regular data audit and model performance evaluation. Historical records are sampled and checked monthly to verify data integrity and logical consistency. Quarterly model evaluation uses cross-validation method to divide the latest data into training set and test set, and compares the difference between predicted fertilizer requirement and actual measured value. Maintenance report records key indicators such as hardware failure rate, algorithm accuracy and user feedback, guiding the direction of subsequent improvement.

[0125] Integration with agricultural meteorological system expands the decision-making dimension. Rainfall forecast data is used to adjust the fertilization time recommendation to avoid fertilizer loss; accumulated temperature data is used to assist in judging the nutrient demand stage of crops; drought warning triggers modification of irrigation and fertilization programs. These cross-system linkages are realized through standardized data exchange protocols, using JSON format to package weather impact coefficients and response strategies.

[0126] Long-term data accumulation forms a digital archive of farmland, recording the soil conditions, fertilization operations and yield changes of each grid year by year. Time series analysis reveals the evolution of soil nutrients, such as the cumulative effect of phosphorus or the depletion trend of potassium. These findings are fed back to the risk assessment model, enhancing its predictive ability for long-term changes. Digital archives support retrospective analysis, helping to explain the historical causes of current problems, and providing data support for precision agriculture.

[0127] The generation of the fertilization optimization strategy is based on the comprehensive evaluation of the farmland fertilizer requirement degree quantification value and the farmer operation bias value. The system first divides the farmland into 0.5 mu operation units, and generates a basic fertilization amount matrix for each unit according to its fertilizer requirement degree value. This matrix contains the recommended amounts of nitrogen, phosphorus, and potassium nutrients, with the numerical range verified by an agronomic expert knowledge base to ensure that it meets the nutritional needs of the crop growth stage. The farmer operation bias value acts as a correction factor on the matrix, with bias levels 1-3 corresponding to a 5%-15% reduction in fertilization amount, and levels 4-5 corresponding to a 10%-20% increase. The correction process takes into account the mobility differences of different nutrients, with the adjustment range of nitrogen fertilizer generally larger than that of phosphorus and potassium fertilizer, and the adjustment sensitivity of sandy soil higher than that of clay soil.

[0128] The generated fertilization optimization strategy is stored in GeoJSON format, with each polygon element containing detailed operation parameters. The fertilization location information is accurate to sub-meter level, with the polygon vertices labeled using WGS84 coordinates. The fertilization timing is dynamically set according to the crop phenological stage and weather conditions, with different operation types such as base fertilizer, tillering fertilizer, and ear fertilizer. Each type is associated with a time window and climate constraint condition. The dosage data is recorded in kg / mu, with the pure nutrient content and the converted commercial fertilizer dosage marked, to facilitate direct operation by farmers. The strategy file is distributed to terminal devices through the digital agriculture platform, supporting online update and offline caching modes.

[0129] The implementation of the fertilization machine control system relies on high-precision positioning and real-time monitoring data. The RTK-GNSS receiver provides centimeter-level positioning services, with a horizontal accuracy of ±2 cm in fixed solution state. The soil moisture sensor collects the volume water content data of the top 20 cm at a frequency of 1 Hz, which is transmitted to the controller after temperature compensation and salinity correction. The positioning data and moisture data are aligned by timestamp and matched in space to the nearest operation unit. The system maintains a sliding window buffer, storing the latest 30 seconds of sensor readings, to calculate the dynamic average and eliminate transient fluctuations.

[0130] The fuzzy PID controller is responsible for adjusting the execution parameters of the fertilization machine. The input variables include position deviation and moisture deviation, with position deviation defined as the vertical distance between the actual travel route and the preset operation trajectory, and moisture deviation referring to the percentage difference between the current water content and the target water content. The output variables control the speed of the stepper motor and the output pressure of the air pump, which jointly regulate the discharge rate and uniformity of the fertilizer. The control rule base contains 81 fuzzy condition statements, covering various possible deviation combinations. The defuzzification process uses the area center method, with an output value update period of 200 ms, ensuring that the system response speed keeps up with the mechanical travel speed.

[0131] The specific process of parameter adjustment is a multi-level closed-loop control. The main control loop looks up the target value in the fertilization optimization strategy based on the current position and generates a theoretical control signal. The secondary control loop receives the theoretical signal and superimposes real-time deviation compensation to output the final execution instruction. The feedback loop continuously monitors the actual discharge amount and verifies the execution effect through the readings of the belt scale or flow meter. When the difference between the actual value and the target value exceeds 5%, the parameter self-tuning program is triggered to adjust the proportional, integral, and derivative coefficients again. The control log records the timestamp, input and output values, and environmental parameters of each adjustment for post-analysis and algorithm optimization.

[0132] The human-machine interface design of the system considers both professionalism and ease of use. The touch screen main interface displays three major parts: operation map, real-time parameters, and alarm information. The operation map is superimposed with satellite images, operation boundaries, and completed areas, and different fertilization amounts are distinguished by gradient colors. The real-time parameter area displays key indicators such as discharge rate, travel speed, and pressure value in numerical and dashboard forms. Alarm information is displayed according to the severity level, with major faults indicated by red flashing and minor abnormalities indicated by yellow warning. Physical buttons retain key functions such as emergency stop and manual speed adjustment to ensure safe operation in case of touch screen failure.

[0133] Safety protection mechanisms are integrated throughout the control process. Mechanical limit switches prevent the hydraulic system from operating beyond its stroke, and electronic overload protection circuits monitor motor current anomalies. Multiple checks are implemented at the software level, including sudden change in fertilization amount, trajectory deviation alarm, and sensor failure detection. When the system detects an abnormal state lasting more than 10 seconds, it automatically enters safety mode, gradually reducing the discharge rate and prompting the operator to intervene. All safety events are recorded in non-volatile memory, including timestamp, device status, and disposal measures, supporting USB export analysis.

[0134] System maintenance functions support remote diagnosis and parameter configuration. Through a 4G network connection to the cloud platform, agricultural experts can view real-time operation data, adjust control parameters, or update fertilization strategies. The device self-check program automatically runs at the first start of the day to check GNSS signal quality, sensor calibration status, and mechanical component wear. Firmware upgrades use differential update technology, downloading only the changed parts to reduce traffic consumption. The maintenance log records the time, content, and operator of each service, forming a complete device health record.

[0135] Field adaptability measures consider various complex operating environments. In areas with poor signal, the system automatically switches to inertial navigation assisted positioning, using gyroscope and odometer data to maintain short-term positioning accuracy. When encountering a head turn, it intelligently identifies the operation boundary and adjusts the discharge mechanism in advance to avoid over-fertilization or missed fertilization at the head. For different physical properties of fertilizers, multiple material parameter presets are built in, covering granular fertilizer, powdery fertilizer, and liquid fertilizer, each type matching corresponding pneumatic conveying and mechanical transmission parameters.

[0136] Integration with farm management systems enables end-to-end data flow. Field information and fertilization prescriptions are automatically downloaded before the start of an operation, and execution data is uploaded in real time during the operation. Performance reports containing operation area, total fertilizer amount, and energy consumption statistics are generated after the operation. The data interface is compatible with the ISOBUS standard, supporting connection with different brands of agricultural machinery. Third-party systems can access fertilization data through RESTful API, but need to be authenticated through OAuth2.0 and subject to rate limits.

[0137] Analysis and mining of long-term operation data form knowledge accumulation. The system automatically identifies high-frequency adjustment areas and marks them as potential problem areas for focused monitoring. It also analyzes operator intervention records to identify areas for improvement in control logic. These findings are integrated into the system through annual software updates, forming a virtuous cycle of continuous improvement.

[0138] In terms of hardware design, industrial-grade components are used to ensure reliability. The control cabinet meets IP65 protection standards, resisting dust and water spray erosion. Electrical components have a working temperature range of -30°C to 70°C, suitable for various climate conditions. Shock-absorbing mounts buffer the impact of field bumps on electronic equipment. The backup power system maintains the operation of key sensors and positioning modules when the main power is off, ensuring data integrity.

[0139] The user training system includes multiple levels. Basic training covers equipment operation and daily maintenance, taught through a combination of on-site demonstrations and video tutorials. Advanced training is aimed at agronomists, explaining system algorithm principles and parameter adjustment methods. Certified engineer training delves into hardware maintenance and software debugging, and upon passing the examination, the engineer obtains technical service qualifications. Training materials are updated with system version updates, and the online knowledge base provides the latest technical documents and common problem solutions.

[0140] Cost control measures make the system feasible for promotion. A general industrial computer is used as the main controller, reducing hardware customization costs. An open-source geographic information system framework is used to process spatial data, reducing software licensing fees. Modular design allows for phased implementation, allowing farmers to first deploy core functions and then gradually expand. Energy efficiency optimization algorithms reduce power consumption, extend device battery life, and reduce energy costs. These designs ensure performance while maintaining a reasonable input-output ratio.

[0141] It is to be understood that the terminology used herein is for the purpose of describing particular embodiments only and is not intended to be limiting; it is not intended to exclude myriad other embodiments of the present application that other inventors can develop based on the same general inventive concepts embodied by the described embodiments. That is, although the present application is described in terms of particular embodiments and illustrative figures, it should be apparent that the scope of the present application is not limited to these specific embodiments.

[0142] While the embodiments of the application have been shown and described herein, it will be understood by those skilled in the art that many changes, modifications, substitutions and alterations to these embodiments can be made without departing from the principles and spirits of the application, and it is intended that the scope of the application be limited solely by the scope of the appended claims and the equivalents thereof.

Claims

1. A method for analyzing and diagnosing farmland fertilizer application based on machine learning, characterized in that, Includes the following steps: Near-infrared spectral data and surface microscopic image data of farmland soil were collected to generate comprehensive soil risk areas; Analyze the gradient of nitrogen migration rate changes within the comprehensive soil risk area and assess the shift in nitrogen migration rate. The interaction intensity between phosphorus and potassium elements within the comprehensive soil risk area was analyzed, and the asymmetric coupling state of phosphorus-potassium interaction was assessed. The fertilizer requirement of farmland is calculated based on the evaluation results of the nitrogen migration rate shift state and the asymmetric coupling state of the phosphorus-potassium interaction. Based on historical fertilization operation records, the matching deviation between farmers' operation behavior and the fertilizer requirements of the farmland is analyzed to generate farmers' operation deviation values. A fertilization optimization strategy is generated by combining the farmland's fertilizer requirements with the farmers' operational deviations. Based on the fertilization optimization strategy, the execution parameters of the fertilizer applicator are dynamically adjusted through an adaptive control algorithm.

2. The method for analyzing and diagnosing farmland fertilizer application based on machine learning according to claim 1, characterized in that, The generated comprehensive soil risk area includes: The near-infrared spectral data were processed using a random forest regression algorithm to generate soil nutrient content distribution information; Soil particle structure features are extracted from the surface microscopic image data using a convolutional neural network model to generate a soil structure feature map. Based on the soil nutrient content distribution information and the soil structure feature map, a comprehensive soil risk area is generated through spatial overlay analysis.

3. The method for analyzing and diagnosing farmland fertilizer application based on machine learning according to claim 2, characterized in that, The process of using a random forest regression algorithm to process the near-infrared spectral data and generate soil nutrient content distribution information specifically includes: The near-infrared spectral data is preprocessed to extract the effective band reflectance data; The effective band reflectance data were trained and modeled using a random forest regression algorithm to predict the content of organic matter, nitrogen, phosphorus, and potassium elements in the soil. Based on the predicted soil element content values, spatial distribution information of soil nutrient content is generated.

4. The method for analyzing and diagnosing farmland fertilizer application based on machine learning according to claim 2, characterized in that, The step of extracting soil particle structure features from the surface microscopic image data using a convolutional neural network model to generate a soil structure feature map specifically includes: The surface microscopic image data is subjected to noise reduction and contrast enhancement processing; The porosity, aggregate morphology, and alignment characteristics of soil particles were extracted using a pre-trained convolutional neural network model. Soil structure feature maps of the farmland area are generated based on the porosity, aggregate morphology, and arrangement direction characteristics.

5. The method for analyzing and diagnosing farmland fertilizer application based on machine learning according to claim 2, characterized in that, The generation of comprehensive soil risk areas based on the soil nutrient content distribution information and the soil structure feature map through spatial overlay analysis specifically includes: Identify the spatial coordinates of areas with abnormal element content in the soil nutrient content distribution information; Identify the spatial coordinates of structural defect regions in the soil structure feature map; Spatial overlay calculations are performed on the regions with abnormal element content and the regions with structural defects to extract the boundaries of the overlapping regions and generate comprehensive soil risk areas.

6. The method for analyzing and diagnosing farmland fertilizer application based on machine learning according to claim 1, characterized in that, The analysis of the gradient of nitrogen migration rate changes within the comprehensive soil risk area, and the assessment of the nitrogen migration rate shift, specifically includes: Obtain nitrogen concentration monitoring data at multiple time points within the comprehensive soil risk area; Calculate the gradient change in nitrogen concentration at adjacent time points to generate a nitrogen migration rate gradient sequence; By comparing the nitrogen migration rate gradient sequence with a preset threshold range, it is determined whether the nitrogen migration rate offset is normal.

7. The method for analyzing and diagnosing farmland fertilizer application based on machine learning according to claim 6, characterized in that, The analysis of the interaction intensity between phosphorus and potassium elements within the comprehensive soil risk area, and the assessment of the asymmetric coupling state of phosphorus-potassium interaction, specifically includes: Spectral response data of phosphorus and potassium elements were collected simultaneously within the comprehensive soil risk area. The correlation strength between the spectral responses of phosphorus and potassium was calculated using the mutual information algorithm. Based on the directional differences in the correlation strength, we quantify whether the asymmetric coupling state of the phosphorus-potassium interaction is reasonable.

8. The method for analyzing and diagnosing farmland fertilizer application based on machine learning according to claim 7, characterized in that, The calculation of farmland nutrient requirements based on the assessment results of the nitrogen migration rate shift state and the asymmetric coupling state of phosphorus-potassium interaction specifically includes: When the deviation of nitrogen migration rate is abnormal, mark the first risk weight coefficient; When the asymmetric coupling state of phosphorus-potassium interaction is unreasonable, mark the second risk weight coefficient; The first risk weight coefficient and the second risk weight coefficient are weighted and fused to output a quantitative value of the fertilizer requirement of farmland.

9. The method for analyzing and diagnosing farmland fertilizer application based on machine learning according to claim 8, characterized in that, The process of analyzing the mismatch between farmers' operational behavior and the fertilizer requirements of the farmland based on historical fertilization operation records to generate farmer operation deviation values ​​specifically includes: Extract time, location, and fertilizer application data from historical fertilization operation records; Compare the quantitative values ​​of the fertilizer requirement of farmland at the same location with the actual fertilizer application data; Calculate the ratio of the standard deviation between the required fertilizer amount and the actual fertilizer amount to generate the farmer's operational deviation value.

10. The method for analyzing and diagnosing farmland fertilizer application based on machine learning according to claim 9, characterized in that, The process of generating a fertilization optimization strategy by combining the farmland's fertilization requirements with the farmer's operational deviations specifically includes: A basic fertilizer application matrix is ​​generated based on the quantified value of the farmland's fertilizer requirement. The farmer's operational deviation value is used to correct the basic fertilizer application matrix; The output includes a fertilization optimization strategy that includes the location, timing, and amount of fertilizer application. The step of dynamically adjusting the execution parameters of the fertilizer applicator according to the fertilization optimization strategy through an adaptive control algorithm specifically includes: The system acquires real-time data on the fertilizer applicator's position and soil moisture, matches the target fertilizer application rate for the current position based on the fertilizer optimization strategy, and dynamically adjusts the fertilizer applicator's feeding rate and spraying pressure parameters using a fuzzy PID controller.

Images