Farmland precise fertilization control method based on multi-source data fusion and deep learning

Abstract

The invention discloses a farmland precise fertilization control method based on multi-source data fusion and deep learning, and the method comprises the steps: obtaining the multi-dimensional heterogeneous data of a target farmland in real time, including the crop spectrum time series data of a satellite image, the ion concentration data collected by a soil sensor network, and the microenvironment time series monitored by a meteorological station; and constructing a Delaune triangulation network spatial index, and re-sampling satellite image data to be matched with a soil sensor network space. Calculating a time-varying mutual information entropy, extracting related microenvironment parameters as coupling factors, combining the coupling factors with soil basic parameters to form a three-dimensional feature matrix, and converting the three-dimensional feature matrix into a growth period feature vector fused with time-space correlation; the collaborative decision model is input, the main branch predicts the basic fertilization amount, the auxiliary branch detects the ion concentration space mutation area and corrects the fertilization amount, a fertilization control instruction is output, and the fertilization device is controlled to conduct precise fertilization. The problem that fertilization control cannot be accurately and efficiently performed on farmland in the prior art is effectively solved.

Description

Technical Field

[0001] This invention relates to the field of fertilization control, and in particular to a method for precision fertilization control of farmland based on multi-source data fusion and deep learning. Background Technology

[0002] With the continuous growth of the global population and the ongoing modernization of agriculture, improving crop yield and quality has become an important goal of agricultural production. Fertilization, as a crucial link in agricultural production, plays a vital role in ensuring crop growth and increasing yield. However, traditional fertilization methods are mostly based on experience and lack scientific and precise guidance, leading to increasingly prominent problems such as inaccurate fertilizer application, low nutrient utilization, and environmental pollution.

[0003] In recent years, the continuous development of sensor technology, satellite remote sensing technology, and big data analysis has provided new ideas and methods for precision fertilization in farmland. By acquiring multi-source data from farmland, such as soil nutrient data, crop growth status data, and meteorological data, and utilizing data fusion and analysis techniques, it is possible to accurately assess and predict farmland fertilization needs, thereby improving the scientific nature and precision of fertilization.

[0004] However, existing precision fertilization technologies still have some shortcomings. On the one hand, the methods for integrating and analyzing multi-source data are not perfect, resulting in low data utilization efficiency and failing to fully realize the value of the data. On the other hand, most existing fertilization decision-making models are based on a single data source or simple statistical analysis methods, making it difficult to comprehensively consider the impact of multiple factors on fertilization needs, leading to insufficient accuracy and reliability of fertilization decisions. These shortcomings prevent existing technologies from accurately and efficiently controlling fertilization in farmland. Summary of the Invention

[0005] This invention provides a method for precise fertilization control of farmland based on multi-source data fusion and deep learning, in order to solve the problem that existing technologies cannot accurately and efficiently control fertilization in farmland.

[0006] Firstly, this application provides a method for precision fertilization control in farmland based on multi-source data fusion and deep learning, including: Real-time acquisition of multidimensional heterogeneous data of the target farmland; the multidimensional heterogeneous data includes crop spectral time-series data from satellite imagery, ion concentration data collected by soil sensor networks, and microenvironment time series monitored by meteorological stations; A spatial index of the Deloitte triangulation is constructed based on the coordinates of soil sensor nodes. Using the vertices of each triangulation in the Deloitte triangulation as a reference, the crop spectral time-series data of the satellite imagery is resampled to satellite data that spatially matches the soil sensor network using an adaptive resolution interpolation algorithm. For each soil sensor node, the time-varying mutual information entropy of the satellite pixel reflectance sequence and ion concentration data corresponding to the satellite data is calculated; microenvironmental parameters in the microenvironment time series that have a correlation with the rate of change of time-varying mutual information entropy reaching a preset threshold are extracted as coupling factors; the time-varying mutual information entropy, coupling factors and preset soil basic parameters are combined into a three-dimensional feature matrix. The three-dimensional feature matrix is ​​converted into a reproductive period feature vector that integrates spatiotemporal correlation; the reproductive period feature vector is input into a preset collaborative decision model, so that the main branch of the collaborative decision model predicts the basic fertilization amount through a gradient boosting decision tree, and the auxiliary branch detects spatial abrupt changes in ion concentration through a graph convolutional network and ion concentration data to correct the basic fertilization amount and output fertilization control instructions; according to the fertilization control instructions, a preset fertilization device is controlled to fertilize the farmland.

[0007] This application acquires multidimensional heterogeneous data of target farmland in real time, including time-series crop spectral data from satellite imagery, ion concentration data collected by a soil sensor network, and microenvironment time series monitored by meteorological stations, enabling a comprehensive and dynamic reflection of the farmland's real-time status. A Deloitte triangulation spatial index is constructed based on the coordinates of soil sensor nodes, and an adaptive resolution interpolation algorithm is used to resample satellite imagery data to match the spatial structure of the soil sensor network, ensuring accurate spatial correspondence and providing a reliable data foundation for subsequent analysis. Furthermore, the time-varying mutual information entropy between satellite pixel reflectance sequences and ion concentration data is calculated, and strongly correlated microenvironmental parameters are extracted as coupling factors. This effectively captures the dynamic correlation between crop growth and environmental factors, providing a scientific basis for precision fertilization. The time-varying mutual information entropy, coupling factors, and basic soil parameters are combined into a three-dimensional feature matrix and converted into a growth period feature vector incorporating spatiotemporal correlations, further enhancing the data's representational capabilities and decision-making value. By predicting the basal fertilizer application rate and detecting spatial abrupt changes in ion concentration through the main and auxiliary branches of a collaborative decision-making model, the basal fertilizer application rate is corrected. This approach comprehensively considers the influence of multiple factors on fertilizer demand, improving the accuracy and reliability of fertilization decisions. Finally, the fertilization device is precisely controlled according to fertilization control commands, achieving automation and intelligence in farmland fertilization. This effectively improves fertilizer utilization, reduces environmental pollution, and enhances crop yield and quality. This application effectively solves the problem that existing technologies cannot accurately and efficiently control fertilization in farmland.

[0008] Furthermore, the real-time acquisition of multidimensional heterogeneous data of the target farmland includes crop spectral time-series data from satellite imagery, ion concentration data collected by a soil sensor network, and microenvironment time-series data monitored by meteorological stations, specifically: High-resolution satellite images of the target farmland are acquired at a preset first time interval using a preset satellite remote sensing system, and time-series data of the spectral reflectance of the crops are extracted. The time-series data of the spectral reflectance includes the reflectance in the visible light band and the infrared band. Based on a pre-set soil sensor network deployed in the target farmland, the concentration data of nitrogen, phosphorus and potassium ions in the soil are collected in real time. The soil sensor network includes multiple sensor nodes, and each node collects data at a pre-set second time interval. The microenvironment time series data of the target farmland is collected in real time through a pre-set meteorological station monitoring system. The microenvironment time series data includes temperature, humidity, light intensity and wind speed. The meteorological station collects data at a pre-set third time interval.

[0009] This application achieves comprehensive monitoring of farmland conditions by acquiring multidimensional heterogeneous data of the target farmland in real time, including time-series crop spectral data from satellite imagery, ion concentration data collected by a soil sensor network, and time-series microenvironment data monitored by a meteorological station. Specifically, high-resolution satellite imagery is acquired at preset time intervals using a satellite remote sensing system to extract time-series data of crop spectral reflectance, covering reflectance in the visible and infrared bands, which can accurately reflect the growth status and health of crops. Simultaneously, based on a soil sensor network deployed within the farmland, real-time concentration data of nitrogen, phosphorus, and potassium ions in the soil are collected, providing direct evidence for soil fertility assessment. Furthermore, real-time collection of microenvironmental data such as temperature, humidity, light intensity, and wind speed through a meteorological station monitoring system helps analyze the impact of environmental factors on crop growth. The integrated acquisition of these multi-source data not only enriches the dimensions of farmland information but also improves the timeliness and accuracy of the data, providing a solid data foundation for subsequent precision fertilization decisions. This enables more scientific and efficient farmland management, improves fertilizer utilization, reduces environmental pollution, and enhances crop yield and quality, resulting in significant economic and ecological benefits.

[0010] Furthermore, the construction of the Deloitte spatial index based on the soil sensor node coordinates specifically involves: Obtain the latitude and longitude coordinates of each node in the soil sensor network, and convert the latitude and longitude coordinates into plane rectangular coordinates; The transformed Cartesian coordinates are deduplicated to remove duplicate coordinate points; According to the point-by-point insertion method, the deduplicated coordinate points are added to the preset initial triangulation in sequence, and the structure of the initial triangulation is adjusted by the empty circumcircle criterion to generate the Deloitte triangulation. Based on the vertex identifiers, adjacent triangle identifiers, and circumcircles of each triangle in the Delonte triangulation, a spatial index table is constructed. The spatial index table uses the triangle vertex coordinates as the search keywords and associates them with the sensor node identifiers corresponding to the triangle vertex coordinates.

[0011] This application achieves efficient management and rapid retrieval of farmland spatial information by constructing a Deloitte triangulation spatial index based on soil sensor node coordinates. Specifically, the latitude and longitude coordinates of each node in the soil sensor network are first obtained and converted into Cartesian coordinates, ensuring the uniformity and operability of the coordinate data. Next, the converted Cartesian coordinates are deduplicated, eliminating duplicate coordinate points to avoid data redundancy and improve data processing efficiency. Then, the deduplicated coordinate points are sequentially added to a pre-defined initial triangulation using a point-by-point insertion method, and the structure of the initial triangulation is adjusted according to the empty circumcircle criterion to generate the Deloitte triangulation. This process not only ensures the structural stability of the triangulation but also optimizes the organization of spatial data. Finally, a spatial index table is constructed based on the vertex identifiers, adjacent triangle identifiers, and circumcircles of each triangle in the Deloitte triangulation. The triangle vertex coordinates are used as search keywords to associate the sensor node identifiers corresponding to the triangle vertex coordinates. This method of constructing the spatial index table enables rapid location and retrieval of sensor nodes and their data related to specific coordinates in subsequent data queries and analysis, greatly improving the speed and accuracy of data processing. Through this series of steps, this application not only optimizes the management of farmland spatial data, but also provides efficient data support for subsequent precision fertilization decisions, thereby improving the scientific nature and precision of fertilization and having significant practical application value.

[0012] Furthermore, the method of resampling the crop spectral time-series data from satellite imagery to satellite data spatially matched with the soil sensor network, using each vertex of the Deloitte triangulation as a reference and an adaptive resolution interpolation algorithm, specifically involves: Calculate the area of ​​each triangle in the Deloitte triangulation, divide the triangles with areas greater than a preset area threshold into multiple sub-triangles with equal areas, until the areas of all triangles and all sub-triangles are less than or equal to the preset area threshold, and determine the interpolation resolution based on the area of ​​the sub-triangles after division. Using the vertices of the triangulation network as interpolation control points, a distance-weighted method is employed to interpolate the reflectance of each band in the crop spectral time-series data of satellite imagery to the centroid position of each sub-triangle. Calculate the deviation between the interpolation result and the original satellite data at the vertices of the triangulation network. If the deviation is greater than the preset deviation threshold, adjust the weighting coefficients and re-interpolate until the deviation is less than or equal to the set deviation threshold. The reflectance of all interpolation points is integrated according to the spatial distribution range of the soil sensor network to form satellite data that spatially matches the soil sensor network.

[0013] This application employs an adaptive resolution interpolation algorithm to resample the time-series crop spectral data from satellite imagery to satellite data spatially matched with a soil sensor network, achieving precise spatial alignment between satellite imagery data and soil sensor data. First, by calculating the area of ​​each triangle within the Deloitte triangulation, triangles with areas exceeding a preset threshold are divided into multiple sub-triangles of equal area, ensuring that the area of ​​all triangles and sub-triangles is less than or equal to the preset area threshold, thus determining the interpolation resolution. This process allows the interpolation resolution to be dynamically adjusted according to the spatial distribution density of the soil sensor network, improving the spatial matching accuracy of the data. Second, using the vertices of the triangulation as interpolation control points, a distance-weighted method is employed to interpolate the reflectance of each band in the time-series crop spectral data from the satellite imagery to the centroid of each sub-triangle. This distance-weighted interpolation method effectively reduces interpolation errors and improves the accuracy of the interpolation results. Furthermore, by calculating the deviation between the interpolation result and the original satellite data at the vertices of the triangulation, and adjusting the weighting coefficients to re-interpolate when the deviation exceeds a preset threshold, the reliability of the interpolation results is ensured. Finally, the reflectance of all interpolation points is integrated according to the spatial distribution range of the soil sensor network to form satellite data that spatially matches the soil sensor network. This process not only improves the spatial consistency of the data but also provides high-quality data support for subsequent precision fertilization decisions in farmland, thereby enabling more scientific guidance for fertilization operations, improving fertilizer utilization, reducing environmental pollution, and increasing crop yield and quality.

[0014] Furthermore, for each soil sensor node, the time-varying cross-information entropy between the satellite pixel reflectance sequence and the ion concentration data corresponding to the satellite data is calculated as follows: For each soil sensor node, the satellite pixel reflectance sequence and corresponding ion concentration data corresponding to each soil sensor node are extracted and aligned according to the time stamp to form two sets of time-synchronized data sequences; The two sets of data sequences are divided into sliding windows of a preset fixed length to obtain multiple sets of reflectance subsequences and ion concentration subsequences; The first probability distribution of the reflectance subsequence, the second probability distribution of the ion concentration subsequence, and the joint probability distribution of the reflectance subsequence and the ion concentration subsequence were calculated using the probability density estimation method. Based on the first probability distribution, the second probability distribution, and the joint probability distribution, the time-varying mutual information entropy sequence is obtained by accumulating and calculating the mutual information entropy corresponding to each window.

[0015] This application achieves precise quantification of the dynamic correlation between crop growth status and soil nutrient dynamics in farmland by calculating the time-varying mutual information entropy of satellite pixel reflectance sequences and ion concentration data corresponding to satellite data. First, for each soil sensor node, the corresponding satellite pixel reflectance sequence and ion concentration data are extracted and aligned according to time stamps to form two time-synchronized data sequences, ensuring the correspondence of data in the time dimension. Next, the two data sequences are divided into sliding windows of preset fixed lengths, resulting in multiple reflectance subsequences and ion concentration subsequences. The sliding window approach captures the changing characteristics of data at different time scales. Then, the probability density estimation method is used to calculate the first probability distribution of the reflectance subsequence, the second probability distribution of the ion concentration subsequence, and the joint probability distribution of both, providing a foundation for calculating the mutual information entropy. Finally, based on these probability distributions, the mutual information entropy corresponding to each window is calculated by summing them to obtain the time-varying mutual information entropy sequence. This process can not only quantify the dynamic relationship between crop growth status and soil nutrients, but also reflect the changing trend of this relationship over time, providing a scientific basis for subsequent precision fertilization decisions, helping to improve fertilizer utilization efficiency, optimize the crop growth environment, and thus improve crop yield and quality.

[0016] Furthermore, the step of extracting microenvironment parameters from the microenvironment time series that have a correlation with the time-varying mutual information entropy change rate reaching a preset threshold as coupling factors specifically involves: Based on the time-varying mutual information entropy sequence, the change rate sequence of the time-varying mutual information entropy is calculated by dividing the difference between the mutual information entropy at the current time and the previous time by the mutual information entropy at the previous time. Each microenvironment parameter in the microenvironment time series is aligned with the rate of change series according to the time marker. The correlation coefficient method was used to calculate the correlation between each microenvironment parameter and the rate of change sequence. Based on a preset correlation judgment threshold, microenvironmental parameters whose correlation reaches the correlation judgment threshold are extracted as coupling factors.

[0017] This application achieves precise identification and quantification of the crop growth environment by extracting microenvironmental parameters closely related to crop growth status as coupling factors. First, based on a time-varying mutual information entropy sequence, the rate of change of time-varying mutual information entropy is obtained by calculating the difference between the current and previous time moments and dividing it by the previous time moment's mutual information entropy. This process reflects the changing trend of crop growth status over time. Next, each microenvironmental parameter in the time series is aligned with the rate of change sequence according to time markers, ensuring data synchronization in the time dimension. Then, the correlation coefficient calculation method is used to calculate the correlation between each microenvironmental parameter and the rate of change sequence. This step quantifies the degree of association between microenvironmental parameters and changes in crop growth status. Finally, based on a preset correlation threshold, microenvironmental parameters whose correlation reaches the threshold are extracted as coupling factors. This process not only screens out microenvironmental parameters that have a significant impact on crop growth but also provides key environmental information for subsequent precision fertilization decisions, thereby achieving more scientific and precise farmland management, improving fertilizer utilization efficiency, optimizing the crop growth environment, and enhancing crop yield and quality.

[0018] Furthermore, the step of combining the time-varying mutual information entropy, coupling factor, and preset soil basic parameters into a three-dimensional feature matrix specifically involves: The time-varying mutual information entropy sequence is normalized to convert the mutual information entropy values ​​at each time point into a preset numerical range. Each parameter sequence of the coupling factor is processed using a normalization method to obtain a normalized coupling factor sequence; Preset soil basic parameters are obtained, and these parameters are associated with soil sensor nodes according to the coordinates of sampling points to form a soil basic parameter sequence. This sequence is then normalized. The soil basic parameters include soil pH, organic matter content, and bulk density. The normalized time-varying mutual information entropy sequence, coupling factor sequence, and soil basic parameter sequence are concatenated in dimensional order to form a three-dimensional feature matrix.

[0019] This application achieves efficient integration and characterization of multi-source farmland data by combining time-varying mutual information entropy, coupling factors, and preset soil basic parameters into a three-dimensional feature matrix. First, the time-varying mutual information entropy sequence is normalized, converting the mutual information entropy values ​​at each time point to a preset numerical range, ensuring data comparability and consistency. Next, each parameter sequence of the coupling factor is processed using a normalization method to obtain a normalized coupling factor sequence, further unifying the data scale. Then, preset soil basic parameters, including soil pH, organic matter content, and bulk density, are obtained and associated with soil sensor nodes according to sampling point coordinates to form a soil basic parameter sequence, which is then normalized. This series of normalization operations not only eliminates dimensional differences between different parameters but also improves data compatibility and stability. Finally, the normalized time-varying mutual information entropy sequence, coupling factor sequence, and soil basic parameter sequence are concatenated in dimensional order to form a three-dimensional feature matrix. This multi-dimensional data integration method can comprehensively reflect the spatiotemporal characteristics and soil properties of farmland, providing high-quality, high-dimensional input data for subsequent deep learning models, thereby improving the model's prediction accuracy and decision-making ability.

[0020] Furthermore, the conversion of the three-dimensional feature matrix into a reproductive period feature vector incorporating spatiotemporal correlation specifically involves: The three-dimensional feature matrix is ​​extracted using a sliding window based on the time dimension to obtain multiple sub-matrices within the time window; the window length of the sliding window is the proportion of the crop growth period, and the window movement step size is half of the window length. For each submatrix, the average value of each feature is calculated along the dimension of the soil sensor node to generate a time-feature two-dimensional matrix; Principal component analysis is used to reduce the dimensionality of the time-feature two-dimensional matrix, retaining principal components whose cumulative contribution rate reaches a preset standard. The principal components after dimensionality reduction are spliced ​​together in chronological order to construct a reproductive period feature vector that integrates spatiotemporal correlation.

[0021] This application achieves efficient dimensionality reduction and feature extraction from multi-source farmland data by converting a three-dimensional feature matrix into a spatiotemporally correlated growth period feature vector. First, a sliding window extraction is performed on the three-dimensional feature matrix according to the time dimension, resulting in multiple sub-matrices within time windows. The window length of the sliding window is the proportion of the crop's growth period, and the window movement step is half the window length. This process captures the characteristic changes of crops at different growth stages, ensuring the continuity and representativeness of the data in the time dimension. Next, for each sub-matrix, the average value of each feature is calculated along the soil sensor node dimension, generating a time-feature two-dimensional matrix. This step further integrates the spatial dimension data and reduces data redundancy. Then, principal component analysis is used to reduce the dimensionality of the time-feature two-dimensional matrix, retaining principal components whose cumulative contribution rate reaches a preset standard. This process not only extracts key information from the data but also reduces data complexity and improves the efficiency of subsequent processing. Finally, the dimensionality-reduced principal components are concatenated in chronological order to construct a growth period feature vector that integrates spatiotemporal correlation. This feature vector, which integrates spatiotemporal information, can more accurately reflect the growth status of crops at different growth stages, providing a more representative and relevant feature representation for precision fertilization decisions. This improves the scientific nature and accuracy of fertilization decisions, helps to increase crop yield and quality, and reduces fertilizer waste and environmental pollution.

[0022] Furthermore, by utilizing the gradient boosting decision tree algorithm in the main branch of the collaborative decision-making model to predict the basic fertilizer application rate, accurate estimation of farmland fertilizer demand is achieved. Specifically, the fertility period feature vector, which incorporates spatiotemporal information, is input into the model's main branch. Using parameters preset in the gradient boosting decision tree algorithm, such as the number of decision trees, the maximum depth of each tree, and the learning step size, the input feature vector is predicted level by level. This process, through the collaborative work of multiple decision trees, gradually optimizes the prediction results, ultimately fusing the calculation results of all decision trees to output an accurate basic fertilizer application rate. This method not only fully utilizes the rich information in the fertility period feature vector but also improves the accuracy and stability of the prediction through level-by-level model optimization. Compared to traditional fertilization methods, this deep learning-based prediction approach can more scientifically guide fertilization operations, avoiding crop growth problems and resource waste caused by improper fertilizer application, thereby effectively improving fertilizer utilization efficiency, promoting healthy crop growth, and reducing negative environmental impacts, resulting in significant economic and environmental benefits.

[0023] Furthermore, the auxiliary branch uses a graph convolutional network and ion concentration data to detect spatial abrupt changes in ion concentration, and corrects the basic fertilization amount accordingly. Specifically: Soil sensor nodes are used as nodes in the graph, with node characteristics being ion concentration data. The graph connection structure is constructed according to the adjacency relationship of the Deloitte triangulation. The graph connection structure is input into the graph convolutional network so that the graph convolutional network outputs the probability that each node belongs to a spatially abrupt region of ion concentration change. Nodes whose probabilities exceed a preset probability judgment threshold are marked as mutation region nodes. Based on the spatial topology of the Deloitte triangulation, adjacent mutation region nodes and the triangular regions where the mutation region nodes are located are merged to form continuous spatial mutation regions of ion concentration. Calculate the difference between the average ion concentration in the region of spatial abrupt change and the average ion concentration in the region of non-abrupt change; Based on the ratio between the difference and the preset standard concentration difference, the basic fertilizer application amount is adjusted, and the adjusted fertilizer application amount is used as the fertilizer control instruction.

[0024] This application utilizes a graph convolutional network and three-dimensional ion concentration data to detect spatially abrupt changes in ion concentration, and adjusts the basic fertilization amount accordingly, achieving refined adjustment of farmland fertilization schemes. First, soil sensor nodes are used as nodes in a graph, with ion concentration data as their feature. A connection structure is constructed based on the adjacency relationship of a Deloitte triangulation, accurately reflecting the spatial relationships between soil sensor nodes. Next, the graph connection structure is input into a graph convolutional network, which outputs the probability that each node belongs to a spatially abrupt change in ion concentration. This step identifies areas of abnormal ion concentration changes in the soil. Then, nodes with probabilities exceeding a preset threshold are marked as abrupt change region nodes. Based on the spatial topology of the Deloitte triangulation, adjacent abrupt change region nodes and their corresponding triangular regions are merged to form continuous spatially abrupt change regions of ion concentration. This process accurately defines the scope of the abrupt change regions. Further, the difference between the average ion concentration in the spatially abrupt change regions and the average ion concentration in the non-abrupt change regions is calculated. Based on the ratio of this difference to the difference in a preset standard concentration, the basic fertilization amount is adjusted, and the adjusted fertilization amount is ultimately used as the fertilization control instruction. This method of adjusting fertilizer application based on the spatial distribution characteristics of ion concentration can accurately correct the uneven distribution of nutrients in the soil, avoid the impact of excessive or insufficient nutrients on crop growth, thereby improving fertilizer utilization efficiency, optimizing crop growth conditions, increasing crop yield and quality, and reducing fertilizer waste and environmental pollution, thus having significant economic and ecological benefits. Attached Figure Description

[0025] Figure 1 This is a flowchart illustrating a method for precision fertilization control of farmland based on multi-source data fusion and deep learning, as provided in this application. Detailed Implementation

[0026] 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.

[0027] Example 1 Please refer to Figure 1 In order to solve the problem that existing technologies cannot accurately and efficiently control fertilization in farmland, this invention provides a method for precise fertilization control of farmland based on multi-source data fusion and deep learning, including steps S01-S04.

[0028] S01: Real-time acquisition of multidimensional heterogeneous data of the target farmland; the multidimensional heterogeneous data includes crop spectral time series data from satellite imagery, ion concentration data collected by soil sensor network, and microenvironment time series monitored by meteorological station.

[0029] As a preferred embodiment of this invention, the real-time acquisition of multidimensional heterogeneous data of the target farmland specifically includes: The time-series data of crop spectral data from satellite imagery is obtained by acquiring multispectral images of the target farmland at a frequency of 1-3 days per acquisition (the preset first time interval) using high-resolution satellites (such as Gaofen-7) or UAV remote sensing systems. As long as the time interval between two adjacent operations is within the range of 1 to 3 days, it meets the frequency requirement. Reflectance data in the 400-900nm band (including the visible light band reflecting crop chlorophyll content and the near-infrared band reflecting biomass) is extracted and sorted according to the acquisition time to form a time-series sequence, which is used to capture the dynamic changes of crop growth stages (such as the jointing stage and the grain-filling stage). The ion concentration data of the soil sensor network is obtained by arranging soil sensor nodes in a 50m × 50m grid (depth 0-30cm, covering the main distribution layer of crop roots) in the target farmland. Each node detects nitrogen (NH4) in the soil in real time through an ion-selective electrode. + NO3 - ), phosphorus (PO4) 3- ), potassium (K) +The concentration of ions, specifically nitrogen, phosphorus, and potassium ions, is uploaded at a frequency of 1-2 hours per upload (a preset second time interval). For example, one sensor might upload data once in the first hour, followed by another upload after a 1.5-hour interval; another sensor might upload data at a fixed interval of 2 hours. As long as the time interval between two consecutive uploads is within the range of 1 to 2 hours, it meets the frequency requirement. This forms a dataset that reflects the spatial heterogeneity of soil nutrients (such as the difference in fertility between the edge and center of a field) and its temporal dynamics (such as the nutrient decay curve after fertilization). Automatic weather stations are set up in farmland to collect parameters such as air temperature, relative humidity, light intensity, and rainfall at a frequency of 15-30 minutes per time (the preset third time interval), forming a continuous record of microenvironmental changes. This record is used to capture the impact of short-term environmental fluctuations (such as diurnal temperature range and precipitation events) on crop nutrient absorption. This application addresses the difficulties in existing technologies caused by the one-sidedness of single-dimensional data evaluation, insufficient timeliness of static data, and poor spatial representativeness leading to the misconception of uniform fertilization, through the collaborative acquisition of multi-dimensional heterogeneous data. Specifically, it manifests itself in the following ways: First, traditional fertilization relies solely on single-point soil sampling (reflecting soil nutrients) or manual observation of crop growth (subjective and lagging), neglecting the synergistic relationship between "crop-soil-environment." For example, if soil nitrogen content is sufficient but low temperature hinders crop absorption, relying solely on soil data may lead to a misjudgment that "fertilization is unnecessary." This solution integrates spectral data (crop status), ion concentration data (soil supply), and microenvironment data (environmental impact) to construct a three-dimensional assessment system, enabling a comprehensive judgment of fertilization needs and avoiding decision-making biases caused by single data sources. Secondly, traditional soil testing relies on laboratory analysis, which takes 3-7 days and cannot cope with the fluctuations in nutrient requirements during the rapid growth period of crops (such as the heading stage of rice and the grain-filling stage of wheat). This solution adopts a "real-time acquisition + time-series analysis" design: a sensor network captures short-term changes in soil ion concentration at high frequency (such as nitrogen leaching after rain), satellite spectral time-series data tracks the dynamic trend of crop growth, and meteorological data records the real-time impact of environmental factors. The combination of these three can promptly detect the dynamic imbalance of crop nutrient supply and demand, providing a basis for dynamically adjusting fertilization plans and avoiding "over-fertilization" or "fertilization deficiency and yield reduction" caused by data lag. Furthermore, farmland soil nutrients exhibit significant spatial heterogeneity (fertility differences can reach 30% within the same field). Traditional single-point sampling can only reflect local conditions, leading to an "average fertilization" pattern (over-fertilization in high-fertility areas results in waste and pollution, while under-fertilization in low-fertility areas affects yield). This solution utilizes a distributed soil sensor network (precisely reflecting local differences) and full-area coverage of satellite imagery (providing a macroscopic understanding of the overall trend of the field), combined with subsequent triangulation spatial indexing to achieve "point-area" data fusion. This provides precise spatial positioning for variable fertilization, solving the resource waste and yield loss problems associated with the traditional "one-size-fits-all" fertilization technique.

[0030] S02: Construct a spatial index of the Deloitte triangulation based on the coordinates of soil sensor nodes; using each vertex of the Deloitte triangulation as a reference, resample the crop spectral time-series data of the satellite imagery to satellite data that spatially matches the soil sensor network through an adaptive resolution interpolation algorithm.

[0031] As a preferred embodiment of this invention, the construction of the Deloitte spatial index based on the soil sensor node coordinates specifically involves: Soil sensors within the target farmland are distributed in a pre-defined grid (e.g., 50m × 50m). Each sensor node records latitude and longitude coordinates, which are first converted to Cartesian coordinates to eliminate geographic projection errors. Then, using all sensor nodes as vertices, a Deloitte triangulation is constructed using the empty circumcircle criterion (i.e., the circumcircle of each triangle does not contain other sensor nodes). This divides the farmland area into a series of continuous, non-overlapping triangles, with each triangle's vertex being a sensor node. Based on this, a spatial index table is constructed, recording the vertex identifiers (corresponding to sensor node IDs), the relationships between adjacent triangles (e.g., triangle numbers sharing sides), and circumcircle parameters (center coordinates, radius). This enables rapid mapping between sensor nodes and spatial regions (e.g., by using a triangle, one can directly query the sensor nodes it contains, or by using a sensor node, locate the triangular region it belongs to). The method involves using the vertices of each triangle in the Deloitte triangulation as a reference, and employing an adaptive resolution interpolation algorithm to resample the crop spectral time-series data from satellite imagery to satellite data spatially matched with the soil sensor network. Specifically: There is an inherent difference between the time-series spectral data of crops from satellite imagery (e.g., 10m resolution) and the spatial scale of soil sensor networks (50m node spacing). Direct fusion would lead to spatial misalignment of the data (e.g., a single satellite pixel covering three sensor nodes, or a single sensor node corresponding to multiple satellite pixels). Therefore, it is necessary to use the vertices of the triangulation network (i.e., sensor node coordinates) as the spatial reference and adjust the spatial scale of the satellite spectral data using an adaptive resolution interpolation algorithm. The adaptive resolution refers to dynamically setting the interpolation accuracy based on the area of ​​the triangles in the triangulation. For triangles with an area greater than a preset threshold (sparsely distributed sensor areas, such as the edge of farmland), a lower resolution interpolation is used (to reduce invalid calculations); for triangles with an area less than the preset threshold (densely distributed sensor areas, such as the center of farmland), a higher resolution interpolation is used (to preserve local details). Taking a farmland with sensors arranged in a 50m×50m grid as an example, if the preset triangle area threshold is 2000㎡ (approximately 45m×45m): Lower resolution (edge ​​triangles with an area > 2000㎡): 10m × 10m resolution interpolation is used, meaning one spectral reflectance estimation point is generated every 10 meters. This results in fewer interpolation points within a single triangle (e.g., a triangle with an area of ​​3000㎡ contains approximately 30 points), reducing invalid calculations caused by sensor sparseness in edge areas and balancing accuracy and efficiency. Higher resolution (central triangle with area < 2000㎡): 5m × 5m resolution interpolation is used, meaning one interpolation point is generated every 5 meters. The number of interpolation points within a single triangle is increased (e.g., a triangle with an area of ​​1000㎡ contains approximately 40 points), which can more finely preserve local details caused by the dense sensor concentration in the central area (such as spectral changes caused by small-scale fertility differences). The core logic of the numerical settings is that resolution is inversely proportional to the area of ​​the triangle (the larger the area, the lower the resolution), and it is necessary to ensure that the interpolation point spacing does not exceed 1 / 5 to 1 / 10 of the sensor grid spacing (e.g., a 50m grid corresponds to a 5-10m resolution) to avoid information loss. Specific values ​​can be flexibly adjusted according to the actual size of the farmland (e.g., 100 mu vs. 1000 mu) and the original accuracy of the spectral data (e.g., 2m resolution for UAV imagery vs. 10m resolution for satellite imagery), with the ultimate goal of "no redundancy in the edge areas and no blurring in the central areas".

[0032] The resampling process uses the spectral values ​​of the vertices of the triangulation network as known control points, and calculates the spectral reflectance inside each triangle (such as the centroid position) using methods such as distance weighting. The specific calculation method for distance weighting is as follows: Based on the spectral reflectance of the three vertices (sensor nodes) of a triangle, the Euclidean distances (d1, d2, d3) from the target point (e.g., the centroid) to the three vertices are calculated using the inverse distance weighted (IDW) formula: Target point reflectance = (r1 / d1 + r2 / d2 + r3 / d3) / (1 / d1 + 1 / d2 + 1 / d3), where r1, r2, and r3 are the spectral reflectance values ​​of the three vertices. The closer the vertex, the greater its contribution to the target point's reflectance. Through this spatial interpolation method, the spectral reflectance at any internal location (e.g., the centroid) can be reasonably calculated based on measured data from the triangle's vertices, achieving a spectral feature mapping from discrete nodes to a continuous region.

[0033] Ultimately, the adjusted satellite data is made to fully match the spatial distribution (node ​​locations, regional divisions) of the sensor network, ensuring the accuracy of the spatial correspondence between "crop spectrum - soil ion concentration" in subsequent analyses; The distance-weighted method is a commonly used data analysis and spatial interpolation method. Its core logic is to estimate unknown data by utilizing the principle that "the closer the known information is, the greater its influence." Specifically, when it is necessary to infer the attribute value of an unknown point based on multiple known data points, this method first identifies the known points surrounding the unknown point. Then, it assigns higher weights to known points closer to the unknown point and lower weights to known points farther away. Finally, it calculates the estimation result of the unknown point using the attribute values ​​of these weighted known points.

[0034] In traditional methods, satellite imagery and soil sensor data often have different spatial resolutions (e.g., 10m / pixel for satellites, 50m / node for sensors). Direct correlation can lead to "one-to-many" or "many-to-one" misalignments (e.g., a satellite pixel covering three sensor nodes may not accurately correspond to the soil data of a single node). In wheat field scenarios, if sensors are densely packed in the eastern part of the field (reflecting local fertility differences) and sparse in the western part (reflecting overall trends), traditional fixed-resolution interpolation can result in the loss of details in the east or data redundancy in the west, ultimately affecting the accuracy of the correlation analysis between "crop spectrum and soil nutrients." This solution uses the spatial partitioning of the Deloitte triangulation and adaptive interpolation to ensure that the spatial scale of satellite data is fully adapted to the sensor network, guaranteeing that the crop spectral data corresponding to each sensor node is accurate and unique, eliminating errors caused by spatial misalignment. Furthermore, spatial correlation of multi-source data relies on global traversal (e.g., querying satellite pixels around a sensor), which can lead to a surge in computation (taking several hours) in large-area farmland (e.g., 1000 mu), failing to meet the needs of real-time fertilization decisions. This solution uses the Deloitte triangulation spatial index to quickly locate the triangular region to which any sensor node belongs and its adjacent nodes, providing efficient support for the subsequent "node-by-node calculation of time-varying mutual information entropy". Furthermore, traditional interpolation methods (such as global fixed resolution) either excessively pursue accuracy (dense interpolation leads to excessive computation) or sacrifice detail (sparse interpolation leads to errors in key areas). For example, in fertilization decisions for cornfields, if the same resolution is used for interpolation at the field edges (sparse sensors) and the center (dense sensors), the edge areas will be slowed down due to redundant computation, while the center area may lose "crop growth mutation points" (such as spectral anomalies caused by local nitrogen deficiency) due to coarse interpolation. This solution's adaptive resolution design achieves "accuracy allocation on demand," ensuring data quality in key areas (dense sensor areas) while reducing computational costs in sparse areas, thus balancing accuracy and efficiency. Finally, a key advantage of the Deloitte triangulation in this step is that its defined adjacency relationships (nodes sharing an edge are adjacent to each other) perfectly align with the intuitive understanding of "spatial proximity" and the first law of geography (things spatially close are more related). The process of constructing the triangulation essentially defines an accurate spatial topology between sensor nodes. This topology is not only the foundation for spatial indexing but also the necessary graph structure input for subsequent Graph Convolutional Network (GCN) analysis (nodes are sensors, and edges are defined by the triangulation's adjacency relationships). Therefore, this step not only serves the current data registration but also builds an indispensable bridge for subsequent intelligent analysis based on spatial relationships (abrupt region detection).

[0035] S03: For each soil sensor node, calculate the time-varying mutual information entropy between the satellite pixel reflectance sequence and ion concentration data corresponding to the satellite data; extract microenvironmental parameters in the microenvironment time series that have a correlation with the rate of change of the time-varying mutual information entropy reaching a preset threshold as coupling factors; combine the time-varying mutual information entropy, coupling factors and preset soil basic parameters into a three-dimensional feature matrix.

[0036] In a preferred embodiment of this invention, the calculation of the time-varying cross-information entropy between the satellite pixel reflectance sequence and the ion concentration data corresponding to each soil sensor node specifically involves: Time-varying mutual information entropy is an indicator that quantifies the dynamic correlation strength between two time-series data (in this case, satellite pixel reflectance sequence and soil ion concentration data): the higher the value, the stronger the correlation between the two (e.g., the higher the synchronicity between crop spectrum changes and soil nitrogen concentration changes). The specific implementation is as follows: 1. For each soil sensor node, extract its corresponding satellite pixel reflectance sequence (such as time-series reflectance in the visible and near-infrared bands) and the soil ion concentration data (time-series data of nitrogen, phosphorus, and potassium concentrations) collected by that node, and align them according to the time stamp (to ensure that the data at the same time point correspond). 2. Divide the two sets of time series data into sliding windows of a fixed length (e.g., 7 days) to obtain multiple subsequences; 3. For each subsequence, the probability distribution of the reflectance subsequence, the probability distribution of the ion concentration subsequence, and the joint probability distribution of the two are calculated by probability density estimation. Then, the mutual information entropy of each window is calculated by summing them up, and finally a time-varying mutual information entropy sequence (reflecting the dynamic change of the correlation strength between crop spectrum and soil nutrients) is formed. Probability density estimation is an algorithm used to estimate the probability distribution characteristics of random variables. Its core objective is to infer the underlying probability density function (i.e., a function describing the likelihood of a random variable occurring around a certain value) from known sample data. Suppose we collect a batch of data (such as nitrogen ion concentration values ​​detected by multiple soil sensors in a field). This data may seem chaotic, but it actually follows a certain distribution pattern. Probability density estimation uses mathematical methods to "reconstruct" this pattern from these samples, such as determining whether these concentration values ​​are more likely to be concentrated in a certain range (e.g., 10-30 mg / kg), or whether they exhibit a uniform distribution or other specific patterns.

[0037] The step of extracting microenvironment parameters from the microenvironment time series that have a correlation with the time-varying mutual information entropy change rate reaching a preset threshold as coupling factors specifically involves: 1. Based on the time-varying mutual information entropy sequence, calculate its rate of change (the difference in mutual information entropy between the current time and the previous time divided by the value of the previous time), and obtain the rate of change sequence; 2. Align the microenvironment time series (temperature, humidity, light, etc.) with the rate of change series by time, and use methods such as Pearson correlation coefficient to calculate the correlation between each microenvironment parameter and the rate of change series; 3. Set a correlation threshold (e.g., |r|≥0.7) and extract parameters that meet the correlation threshold (e.g., light intensity, temperature) as coupling factors (i.e., environmental factors that significantly affect the change in the "crop-soil" correlation strength). The process of combining the time-varying mutual information entropy, coupling factor, and preset soil basic parameters into a three-dimensional feature matrix specifically involves: The time-varying mutual information entropy sequence, coupling factor sequence (such as light intensity and temperature), and basic soil parameters (pH value, organic matter content, and bulk density) were normalized (converted to the [0,1] interval to eliminate dimensional differences); the three types of data were spliced ​​together according to the structure of "time dimension (such as daily / hourly), soil sensor node dimension (spatial distribution), and feature type dimension (mutual information entropy, coupling factor, and basic parameters)" to form a three-dimensional matrix (for example, a certain element in the matrix represents the normalized value of "day 3, sensor node 5, and light intensity"). This approach quantifies the dynamic correlation between crop spectra and soil ions through time-varying mutual information entropy (e.g., increased correlation during the jointing stage indicates high crop dependence on soil nitrogen). This enables the model to identify complex scenarios such as "sufficient soil nutrients but insufficient crop absorption," avoiding decision-making biases caused by single data points. Traditional fertilization models often use soil nutrient data or crop growth data alone, ignoring their dynamic correlation. For example, during the wheat jointing stage, focusing solely on soil nitrogen concentration (static sufficiency) while ignoring the decrease in crop spectral reflectance (reflecting actual nitrogen deficiency) can lead to insufficient fertilization; conversely, focusing solely on the spectrum while ignoring the actual soil nitrogen reserves can lead to over-fertilization. Furthermore, traditional models often feature single-dimensional (e.g., soil nutrients only) or two-dimensional (e.g., time-soil nutrient) features, failing to simultaneously reflect temporal dynamics, spatial distribution, and the correlation of multiple feature types. For instance, in a 100-acre cornfield, the strength of the crop-soil correlation, the degree of environmental influence, and basic soil properties vary across different areas (sensor nodes). Two-dimensional features cannot distinguish these spatial differences, resulting in low accuracy in fertilization decisions for local areas. The three-dimensional feature matrix of this scheme simultaneously covers temporal dynamics (temporal changes), spatial distribution (sensor nodes), and feature types (correlation strength, environmental influence, and basic attributes), providing comprehensive input for subsequent deep learning models and enabling the models to accurately capture the complex patterns of "spatiotemporal heterogeneity + multi-factor correlation".

[0038] S04: Convert the three-dimensional feature matrix into a reproductive period feature vector that integrates spatiotemporal correlation; input the reproductive period feature vector into a preset collaborative decision-making model, so that the main branch of the collaborative decision-making model predicts the basic fertilization amount through a gradient boosting decision tree, and the auxiliary branch detects spatial abrupt changes in ion concentration through a graph convolutional network and ion concentration data, corrects the basic fertilization amount, and outputs fertilization control instructions; according to the fertilization control instructions, control the preset fertilization device to fertilize the farmland.

[0039] In a preferred embodiment of this invention, the step of converting the three-dimensional feature matrix into a reproductive period feature vector that incorporates spatiotemporal correlations specifically involves: For the time dimension, time windows are divided according to crop growth stages (such as the jointing and grain-filling stages of wheat, and the tillering and booting stages of rice). Time-series data within each window are aggregated (such as calculating the mean of time-varying mutual information entropy, the peak value of coupling factors, and the stability index of basic soil parameters) to capture key dynamics within the growth stage. For the spatial dimension, based on the regional division of the Deloitte triangulation, the spatial mean and coefficient of variation of sensor node features within each triangular region are calculated to reflect the spatial consistency and differences within the region. The "temporal aggregation features (such as mutual information entropy trend) + spatial aggregation features (such as regional average ion concentration) + growth stage identifier (such as jointing stage code)" of each growth stage are concatenated into a one-dimensional vector to form a growth stage feature vector, achieving deep integration of "temporal dynamics - spatial distribution - growth stage" (for example, the jointing stage feature vector contains core information such as the correlation strength between crops and soil at that stage, regional nutrient distribution differences, and key environmental impact factors). This solution accurately captures the "temporal dynamics (such as nutrient demand rhythm) + spatial distribution (such as local nutrient deficiency)" of each growth stage by transforming a "three-dimensional feature matrix" into a "growth stage feature vector." This ensures that the feature input is deeply adapted to the growth stage, solving the problem of "mismatch between stage requirements and fertilization plans." Traditional fertilization models often use a "one-size-fits-all" approach to feature input (such as using only the average soil nutrient level), ignoring the spatiotemporal differences in crop requirements at different growth stages. For example, corn is sensitive to phosphorus requirements during the seedling stage (spatial distribution is uniform), while its nitrogen requirements surge during the grain-filling stage (and are greatly affected by local soil fertility). If traditional models use the same set of features, it will lead to excessive phosphorus application during the seedling stage and insufficient nitrogen application during the grain-filling stage.

[0040] The step of inputting the reproductive period feature vector into a preset collaborative decision-making model, so that the main branch of the collaborative decision-making model predicts the basic fertilizer application amount through a gradient boosting decision tree, specifically involves: The growth period feature vector is input into the GBDT model. The model learns the nonlinear relationship between features and fertilizer application by step by step through preset decision tree number (e.g., 100 trees), maximum depth of each tree (e.g., 8 layers) and learning step size (e.g., 0.1). For example, the higher the mutual information entropy at the jointing stage, the more the crop's nitrogen demand depends on the soil supply, and the basic nitrogen application should be increased accordingly. Finally, it outputs the basic fertilizer application based on the overall growth period features (reflecting the overall nutrient demand baseline of the field). The GBDT (Gradient Boosting Decision Tree) model is a machine learning model based on ensemble learning. It constructs multiple decision trees and combines them to complete prediction tasks. Its core principle is the "gradient boosting" strategy, where each new tree is trained on the error (residual) generated by the previous model's predictions, continuously reducing the error to improve the overall model's performance. Specifically, the model starts with an initial predicted value (such as the mean). At each step, based on the difference between the current prediction and the true value (gradient direction), it generates a decision tree that minimizes the error to the greatest extent. Then, the prediction results of the new tree are added to the previous model with certain weights. This process is repeated until a preset number of trees is reached or the error meets the requirements. GBDT can be used for classification and regression tasks, has the ability to handle non-linear data and automatically capture feature interactions, and demonstrates strong prediction accuracy and robustness in practical applications.

[0041] The auxiliary branch uses a graph convolutional network and ion concentration data to detect spatial abrupt changes in ion concentration, corrects the basic fertilization amount, and outputs fertilization control instructions, specifically: Using soil sensor nodes as graph nodes, with node features being ion concentration data (spatial distribution of nitrogen, phosphorus, and potassium), a graph structure is constructed based on the adjacency relationships of the Deloitte triangulation. The graph structure is input into the GCN, and the model learns the spatial relationships between nodes through multi-layer convolution operations, outputting the probability that each node belongs to a "spatial abrupt change region of ion concentration" (e.g., if the nitrogen concentration in a certain area suddenly drops below 30% of the surrounding area, the probability is close to 1). Regions with probabilities exceeding a threshold (e.g., 0.8) are marked as abrupt change regions, and the concentration difference between them and non-abrupt change regions is calculated. The base fertilizer application is corrected according to a preset ratio (e.g., for every 10% increase in the difference, the fertilizer application rate is adjusted by 15%) to obtain a precise fertilizer application rate that takes into account both the overall baseline and local anomalies. The GCN (Graph Convolutional Network) is a deep learning model specifically designed for processing graph-structured data. It borrows the ideas of convolutional neural networks, updating node representations by aggregating the features of each node itself with those of its neighboring nodes, thereby capturing the dependencies between nodes in the graph. Its core principle is to utilize the adjacency information and node features of the graph, using convolutional operations to allow each node to "perceive" the information of its surrounding neighbors, gradually learning more globally meaningful node embeddings. Integrate the results from the main and auxiliary branches to generate an instruction containing "fertilizer application amount - spatial coordinates - execution time" (e.g., "Apply 12 kg / mu of nitrogen to the abrupt change area in the eastern part of the field, and apply 8 kg / mu of nitrogen to other areas, to be executed at 9:00 the next day"). In the collaborative decision-making model of this application, the main branch (GBDT) grasps the overall needs of the field (based on the baseline quantity of growth period characteristics), while the auxiliary branch (GCN) accurately identifies local mutation areas (based on the spatial correlation of ion concentration). The combination of the two achieves "global baseline + local correction," solving the problem of "separation between global and local needs." This is because existing technologies either use global models (such as linear regression) to predict fertilizer application, ignoring local soil anomalies (such as nutrient loss in a certain area due to water leakage); or they only use local sampling data, lacking overall trend control. For example, in a 100-mu cotton field, the overall soil potassium content is sufficient (the global model predicts low potassium application), but the potassium concentration in the western part drops sharply due to uneven irrigation (local anomaly). Traditional models would lead to reduced yield in the western part because they ignore local anomalies.

[0042] Furthermore, the fertilization control command is transmitted to a preset variable fertilization device (such as a smart fertilizer spreader equipped with GPS). The device locates the target area according to the spatial coordinates in the command and executes the specified fertilization amount precisely, thus completing closed-loop control.

[0043] To further understand the fertilization method of this application, this application illustrates the fertilization process using a specific scenario: The following section details the entire process of the method using a 150-mu summer maize field (variety "Zhengdan 958") in the North China Plain as an example. The growth period of this maize field is from June to September (approximately 100 days), covering the seedling stage (mid-June to mid-July), jointing stage (mid-July to early August), large trumpet stage (early August to mid-August), and grain-filling stage (mid-August to mid-September). The core requirement is to match the characteristics of maize, such as "short growth period, concentrated fertilizer requirement, and sensitivity to potassium," to achieve precision fertilization.

[0044] Step 1: Multi-source data collection: Constructing a "crop-soil-environment" data system; Crop spectral data: Multispectral images were collected using a drone remote sensing system (every 5 days), with a focus on extracting the red edge reflectance in the 700-750nm band (the higher the nitrogen content in maize leaves, the lower the red edge reflectance), forming time-series data of "date-pixel-band" (e.g., the red edge reflectance is generally higher during the seedling stage, reflecting lower nitrogen content). Soil ion concentration data: 42 soil sensor nodes (depth 0-30cm) were arranged in a 60m×60m grid, and NH4 was collected every 2 hours. + NO3 - K + Concentration, forming a dynamic dataset of "date-node-ion type" (e.g., K during the large trumpet mouth period). + (The concentration decrease rate reaches 0.5 mg / kg / day, which requires close monitoring). Microenvironment data: Field weather stations record temperature (25-38℃), humidity (40%-90%), light intensity (50,000-120,000 lux), and rainfall every 30 minutes to form a time series of "date-environmental factors" (e.g., when the temperature is ≥35℃ during the jointing stage, the nitrogen absorption rate of maize is significantly increased).

[0045] Step 2: Spatial Fusion: Resolving scale differences between multi-source data; Constructing the Deloitte spatial index: Convert the latitude and longitude of 42 sensor nodes into planar coordinates, construct a triangulation network (forming 65 triangular regions) according to the "empty circumcircle criterion", record the vertex (sensor ID) and adjacency relationship of each triangle (such as triangles T23 and T24 sharing node 18), realize the fast mapping of "sensor node - spatial region" (when querying soil data of a certain region, directly associate it with its vertex sensor). Satellite data resampling: The original resolution of the UAV imagery is 2m (single pixel coverage 4㎡), which does not match the 60m spacing of the sensor nodes. Using the vertices of the triangulation network as a reference, a distance-weighted interpolation method is used: a resolution of 10m is used for edge triangles with an area >3000㎡ (sparse sensors), and a resolution of 5m is used for central triangles with an area <800㎡ (dense sensors). This ultimately makes the red-edge band reflectance data spatially match the sensor network (each node corresponds to a unique spectral time sequence).

[0046] Step 3: Feature Extraction: Quantifying the correlation patterns between "crop-soil-environment"; Calculate the time-varying cross-information entropy: For each sensor node, compare the red-edge band reflectivity sequence with NO3. - K + Concentration sequences were aligned by time, and mutual information entropy (reflecting association strength) was calculated using a 5-day sliding window. Results showed that the red-edge reflectance during the large trumpet-shaped mouth period was related to NO3. - The mutual information entropy reached 0.91 (significantly higher than 0.35 at the seedling stage), indicating a strong correlation between maize nitrogen demand and soil supply at this stage; and with K + The mutual information entropy reached 0.88, reflecting the dependence on potassium during the critical period of lodging resistance; Extracting coupling factors: Calculate the rate of change of time-varying mutual information entropy (e.g., a 15% increase on the 3rd day of the large trumpet mouth period compared to the 2nd day), perform correlation analysis with meteorological data, and find that "temperature (r=0.82)" and "light intensity (r=0.76)" are the core coupling factors (under high temperature and strong light, the absorption efficiency of nitrogen and potassium by corn increases by 30%). Construct a three-dimensional feature matrix: incorporating time-varying mutual information entropy, coupling factors (temperature, light), and basic soil parameters (pH=6.8, organic matter 1.8%, bulk density 1.2 g / cm³). 3 After normalization, the data is assembled into a three-dimensional matrix (100 days × 42 nodes × 7 features) according to "time (days) - sensor nodes - feature type" to achieve multi-dimensional information fusion; Step 4: Decision Model: The main and auxiliary branches work together to output precise fertilizer application amounts; Generate a feature vector representing the growth period: Divide the time window according to the maize growth period and aggregate spatiotemporal features on the three-dimensional matrix: Seedling stage: Focus on "low mutual information entropy + stability of basic soil parameters" (e.g., organic matter variation coefficient < 5%). The "big horn mouth" period: emphasizing "high mutual information entropy peak value + temperature / illuminance coupling intensity + K". + Spatial variability coefficient (e.g., regional potassium concentration differences of up to 20% require special correction); This ultimately forms a one-dimensional feature vector for each reproductive stage (e.g., the vector for the large trumpet stage contains 26 dimensions, covering nitrogen-potassium correlation strength, environmental impact weight, etc.).

[0047] Collaborative decision-making model inference: Main branch (gradient boosting decision tree): Input feature vector, learn non-linear relationships through 150 decision trees (e.g., for every 10% decrease in red edge reflectivity during the large trumpet mouth stage, the basic nitrogen application rate increases by 3 kg / mu), and output basic fertilizer application rate (basic nitrogen 15 kg / mu and potassium 8 kg / mu during the large trumpet mouth stage). Auxiliary branch (graph convolutional network): with 42 nodes K + Using concentration as a feature, a graph structure was constructed based on the adjacency relationship of a triangular network, and K values ​​were detected in three nodes (triangle T11) in the northwest region of the field. + The concentration (50 mg / kg) was significantly lower than that of the surrounding area (70 mg / kg), and was marked as a spatial mutation area; the difference was calculated to be 20 mg / kg (the preset standard deviation was 10 mg / kg), and the potassium fertilizer application rate in this area was adjusted to 12 kg / mu proportionally; Output fertilization instructions: Generate a control instruction that reads "Apply 15 kg / mu of nitrogen and 12 kg / mu of potassium in the Northwest mutation area, and apply 15 kg / mu of nitrogen and 8 kg / mu of potassium in other areas, to be executed on August 5th".

[0048] Step 5: Fertilization Implementation and Dynamic Adjustment; After receiving instructions, the intelligent fertilizer applicator (equipped with GPS and a variable-rate fertilization device) navigates according to the coordinates of the triangular area and precisely executes differentiated fertilization. After 10 days, steps 1-4 are repeated, and the fertilization strategy during the grouting period is adjusted based on new data (such as nitrogen concentration returning to a reasonable range after the large trumpet stage) (reducing nitrogen application by 30% and increasing phosphorus application by 10%), forming a dynamic closed loop. During the jointing stage of the aforementioned 150 mu (approximately 10 hectares) summer corn field (late July), it experienced two consecutive days of extreme rainfall (daily rainfall reaching 80 mm), causing a sudden change in microenvironmental parameters (humidity jumped from 55% to 92%, and light intensity dropped from 100,000 lux to 30,000 lux). Simultaneously, NO3 in the soil surface layer... - The concentration decreases by an average of 40% due to leaching (up to 60% in low-lying areas). This application's method achieves dynamic response through the following logic: Step 1: Mutation data capture and labeling; Microenvironment parameter monitoring: When the weather station detects that the 24-hour humidity change rate reaches 67% (exceeding the preset threshold of 50%) and the light intensity decreases by 70% (exceeding the preset threshold of 40%), the system automatically marks it as a "microenvironmental mutation event" and triggers data priority upgrade (subsequent analysis will prioritize the use of the latest data); Soil data anomaly identification: The soil sensor network showed that 12 out of 42 nodes (mainly located in low-lying areas) had NO3... - The concentration dropped from 35 mg / kg to 14 mg / kg within 12 hours after rainfall, and was marked as a "nutrient drop point" using the coefficient of variation method (the deviation of the current value from the historical mean is greater than 3 times the standard deviation).

[0049] Step 2: Dynamic update of the feature layer; Time-varying mutual information entropy reconstruction: Red edge reflectance and NO3 before rainfall - The mutual information entropy of concentration was 0.65 (moderate correlation); after rainfall, due to the obstruction of root absorption, the correlation between the two weakened, and the mutual information entropy dropped to 0.32, with a change rate of -51% (negative values ​​indicate a decrease in correlation strength). Coupling factors were re-screened: through correlation calculation, the correlation between "humidity" and the rate of change of mutual information entropy after rainfall surged from 0.35 to 0.89 (becoming the primary coupling factor), while the correlation of "sunlight" dropped to 0.21 (being excluded), indicating that "high humidity inhibits nitrogen absorption" is the core influencing mechanism at this time; Three-dimensional feature matrix adjustment: The feature values ​​of abrupt change regions are weighted and corrected (e.g., the nitrogen concentration feature weight of low-lying nodes is increased from 0.2 to 0.35) to ensure that the matrix can reflect local anomalies under extreme weather conditions.

[0050] Step 3: Emergency correction of the decision-making model; Main branch (GBDT) base quantity adjustment: Input the updated jointing stage feature vector (high humidity + low nitrogen association + nutrient drop marker), and the model relearns through 150 decision trees: The basic nitrogen application rate has been increased from the original plan of 10 kg / mu to 14 kg / mu (to compensate for leaching losses). Introducing a "post-rainfall delayed absorption coefficient" (value 0.7), the actual effective nitrogen application rate is calculated as 70% (to avoid premature fertilization leading to secondary leaching); Spatial mutation enhancement of auxiliary branches (GCN): For 12 "nutrient drop nodes" in low-lying areas, the probability of the graph convolutional network detecting that they belong to the ion concentration mutation region is 0.93 (far higher than the normal threshold of 0.8). The nitrogen concentration difference between this region and the non-mutation region is calculated to be 22 mg / kg (the preset standard deviation is 10 mg / kg). The nitrogen application rate is further corrected to 18 kg / mu according to the ratio, and "preferred fertilization areas" are marked. Output emergency instructions: Generate a control instruction that “apply nitrogen at 18 kg / mu in low-lying areas with sudden changes, and apply nitrogen at 14 kg / mu in other areas, with the execution time delayed until 48 hours after the end of rainfall (after the soil moisture content drops below 60%)”, along with an additional instruction that “potassium fertilizer should not be applied for the time being” (to avoid potassium ion fixation in high humidity environments). Step 4: Execute feedback and make secondary adjustments; The fertilizer applicator was instructed to perform the operation on July 28 (two days after the rainfall ended). Monitoring data three days later showed: Low-lying areas NO3 - The concentration rose back to 28 mg / kg (close to normal levels), and the red edge reflectance decreased by 15% (reflecting the recovery of nitrogen absorption in corn). The time-varying mutual information entropy rebounded to 0.58 (correlation strength recovered), and the correlation between humidity and the rate of change of mutual information entropy decreased to 0.42 (environmental impact weakened). Based on this, the system judges that the impact of extreme weather has subsided and reverts the fertilization plan for the next stage (the large funnel stage) back to normal logic to avoid overcompensation leading to nutrient excess; Thanks to the emergency decisions made above, the cornfield did not experience "nutrient depletion and yellowing" or "nutrient loss" due to extreme rainfall.

[0051] The specific embodiments described above further illustrate the purpose, technical solution, and beneficial effects of the present invention. It should be understood that the above descriptions are merely specific embodiments of the present invention and are not intended to limit the scope of protection of the present invention. In particular, it should be noted that any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention for those skilled in the art.

Claims

1. A method for precision fertilization control in farmland based on multi-source data fusion and deep learning, characterized in that, include: Real-time acquisition of multidimensional heterogeneous data of the target farmland; The multidimensional heterogeneous data includes crop spectral time-series data from satellite imagery, ion concentration data collected by soil sensor networks, and microenvironment time-series data monitored by meteorological stations. Construct a spatial index for the Deloitte triangulation based on the coordinates of soil sensor nodes; Using the vertices of each triangle in the Deloitte triangulation as a reference, the crop spectral time series data of satellite imagery is resampled to satellite data that spatially matches the soil sensor network through an adaptive resolution interpolation algorithm; For each soil sensor node, the time-varying mutual information entropy of the satellite pixel reflectance sequence and ion concentration data corresponding to the satellite data is calculated; microenvironmental parameters in the microenvironment time series that have a correlation with the rate of change of time-varying mutual information entropy reaching a preset threshold are extracted as coupling factors; the time-varying mutual information entropy, coupling factors and preset soil basic parameters are combined into a three-dimensional feature matrix. The three-dimensional feature matrix is ​​converted into a reproductive period feature vector that integrates spatiotemporal correlation; the reproductive period feature vector is input into a preset collaborative decision model, so that the main branch of the collaborative decision model predicts the basic fertilization amount through a gradient boosting decision tree, and the auxiliary branch detects spatial abrupt changes in ion concentration through a graph convolutional network and ion concentration data to correct the basic fertilization amount and output fertilization control instructions; according to the fertilization control instructions, a preset fertilization device is controlled to fertilize the farmland.

2. The method for precision fertilization control of farmland based on multi-source data fusion and deep learning according to claim 1, characterized in that, The process involves real-time acquisition of multidimensional heterogeneous data of the target farmland. This multidimensional heterogeneous data includes time-series crop spectral data from satellite imagery, ion concentration data collected by a soil sensor network, and time-series microenvironment data monitored by a meteorological station. Specifically: High-resolution satellite images of the target farmland are acquired at a preset first time interval using a preset satellite remote sensing system, and time-series data of the spectral reflectance of the crops are extracted. The time-series data of the spectral reflectance includes the reflectance in the visible light band and the near-infrared band. Based on a pre-set soil sensor network deployed in the target farmland, the concentration data of nitrogen, phosphorus and potassium ions in the soil are collected in real time. The soil sensor network includes multiple sensor nodes, and each node collects data at a pre-set second time interval. The microenvironment time series data of the target farmland is collected in real time through a pre-set meteorological station monitoring system. The microenvironment time series data includes temperature, humidity, light intensity and wind speed. The meteorological station collects data at a pre-set third time interval.

3. The method for precision fertilization control of farmland based on multi-source data fusion and deep learning according to claim 1, characterized in that, The construction of the Deloitte spatial index based on the coordinates of soil sensor nodes is specifically as follows: Obtain the latitude and longitude coordinates of each node in the soil sensor network, and convert the latitude and longitude coordinates into plane rectangular coordinates; The transformed Cartesian coordinates are deduplicated to remove duplicate coordinate points; According to the point-by-point insertion method, the deduplicated coordinate points are added to the preset initial triangulation in sequence, and the structure of the initial triangulation is adjusted by the empty circumcircle criterion to generate the Deloitte triangulation. Based on the vertex identifiers, adjacent triangle identifiers, and circumcircles of each triangle in the Delonte triangulation, a spatial index table is constructed. The spatial index table uses the triangle vertex coordinates as the search keywords and associates them with the sensor node identifiers corresponding to the triangle vertex coordinates.

4. The method for precision fertilization control of farmland based on multi-source data fusion and deep learning according to claim 1, characterized in that, The method involves using the vertices of each triangle in the Deloitte triangulation as a reference, and employing an adaptive resolution interpolation algorithm to resample the crop spectral time-series data from satellite imagery to satellite data spatially matched with the soil sensor network. Specifically: Calculate the area of ​​each triangle in the Deloitte triangulation, divide the triangles with areas greater than a preset area threshold into multiple sub-triangles with equal areas, until the areas of all triangles and all sub-triangles are less than or equal to the preset area threshold, and determine the interpolation resolution based on the area of ​​the sub-triangles after division. Using the vertices of the triangulation network as interpolation control points, a distance-weighted method is employed to interpolate the reflectance of each band in the crop spectral time-series data of satellite imagery to the centroid position of each sub-triangle. Calculate the deviation between the interpolation result and the original satellite data at the vertices of the triangulation network. If the deviation is greater than the preset deviation threshold, adjust the weighting coefficients and re-interpolate until the deviation is less than or equal to the set deviation threshold. The reflectance of all interpolation points is integrated according to the spatial distribution range of the soil sensor network to form satellite data that spatially matches the soil sensor network.

5. The method for precision fertilization control of farmland based on multi-source data fusion and deep learning according to claim 1, characterized in that, For each soil sensor node, the time-varying cross-information entropy between the satellite pixel reflectance sequence and the ion concentration data corresponding to the satellite data is calculated as follows: For each soil sensor node, the satellite pixel reflectance sequence and corresponding ion concentration data corresponding to each soil sensor node are extracted and aligned according to the time stamp to form two sets of time-synchronized data sequences; The two sets of data sequences are divided into sliding windows of a preset fixed length to obtain multiple sets of reflectance subsequences and ion concentration subsequences; The first probability distribution of the reflectance subsequence, the second probability distribution of the ion concentration subsequence, and the joint probability distribution of the reflectance subsequence and the ion concentration subsequence were calculated using the probability density estimation method. Based on the first probability distribution, the second probability distribution, and the joint probability distribution, the time-varying mutual information entropy sequence is obtained by accumulating and calculating the mutual information entropy corresponding to each window.

6. The method for precision fertilization control of farmland based on multi-source data fusion and deep learning according to claim 1, characterized in that, The step of extracting microenvironment parameters from the microenvironment time series that have a correlation with the time-varying mutual information entropy change rate reaching a preset threshold as coupling factors specifically involves: Based on the time-varying mutual information entropy sequence, the change rate sequence of the time-varying mutual information entropy is calculated by dividing the difference between the mutual information entropy at the current time and the previous time by the mutual information entropy at the previous time. Each microenvironment parameter in the microenvironment time series is aligned with the rate of change series according to the time marker. The correlation coefficient method was used to calculate the correlation between each microenvironment parameter and the rate of change sequence. Based on a preset correlation judgment threshold, microenvironmental parameters whose correlation reaches the correlation judgment threshold are extracted as coupling factors.

7. The method for precision fertilization control of farmland based on multi-source data fusion and deep learning according to claim 1, characterized in that, The process of combining the time-varying mutual information entropy, coupling factor, and preset soil basic parameters into a three-dimensional feature matrix specifically involves: The time-varying mutual information entropy sequence is normalized to convert the mutual information entropy values ​​at each time point into a preset numerical range. Each parameter sequence of the coupling factor is processed using a normalization method to obtain a normalized coupling factor sequence; Preset soil basic parameters are obtained, and the soil basic parameters are associated with soil sensor nodes according to the sampling point coordinates to form a soil basic parameter sequence. The soil basic parameter sequence is then normalized. The soil basic parameters include soil pH value, organic matter content, and bulk density. The normalized time-varying mutual information entropy sequence, coupling factor sequence, and soil basic parameter sequence are concatenated in dimensional order to form a three-dimensional feature matrix.

8. The method for precision fertilization control of farmland based on multi-source data fusion and deep learning according to claim 1, characterized in that, The process of converting the three-dimensional feature matrix into a reproductive period feature vector that incorporates spatiotemporal correlations specifically involves: The three-dimensional feature matrix is ​​extracted using a sliding window based on the time dimension to obtain multiple sub-matrices within the time window; the window length of the sliding window is the proportion of the crop growth period, and the window movement step size is half of the window length. For each submatrix, the average value of each feature is calculated along the dimension of the soil sensor node to generate a time-feature two-dimensional matrix; Principal component analysis is used to reduce the dimensionality of the time-feature two-dimensional matrix, retaining principal components whose cumulative contribution rate reaches a preset standard. The principal components after dimensionality reduction are spliced ​​together in chronological order to construct a reproductive period feature vector that integrates spatiotemporal correlation.

9. The method for precision fertilization control of farmland based on multi-source data fusion and deep learning according to claim 1, characterized in that, The main branch of the collaborative decision-making model predicts the basic fertilizer application amount through gradient boosting of the decision tree, specifically: The reproductive period feature vector is input into the main branch of the collaborative decision-making model, so that the main branch of the collaborative decision-making model predicts the input reproductive period feature vector step by step by increasing the number of decision trees, the maximum depth of each tree and the learning step size in the gradient boosting decision tree algorithm, and integrates the calculation results of all decision trees to output the basic fertilization amount.

10. The method for precision fertilization control of farmland based on multi-source data fusion and deep learning according to claim 1, characterized in that, The auxiliary branch uses a graph convolutional network and ion concentration data to detect spatial abrupt changes in ion concentration and corrects the basic fertilization amount, specifically as follows: Soil sensor nodes are used as nodes in the graph, with node characteristics being ion concentration data. The graph connection structure is constructed according to the adjacency relationship of the Deloitte triangulation. The graph connection structure is input into the graph convolutional network so that the graph convolutional network outputs the probability that each node belongs to a spatially abrupt region of ion concentration change. Nodes whose probabilities exceed a preset probability judgment threshold are marked as mutation region nodes. Based on the spatial topology of the Deloitte triangulation, adjacent mutation region nodes and the triangular regions where the mutation region nodes are located are merged to form continuous spatial mutation regions of ion concentration. Calculate the difference between the average ion concentration in the region of spatial abrupt change and the average ion concentration in the region of non-abrupt change; Based on the ratio between the difference and the preset standard concentration difference, the basic fertilizer application amount is adjusted, and the adjusted fertilizer application amount is used as the fertilizer control instruction.

Images