A method and device for precision variable fertilization of farmland soil by fusing multi-modal data

Abstract

The application discloses a kind of farmland soil precision variable fertilization method and device of fusion multi-modal data, comprising: fusion with the soil, topography etc. Multisource data collected with slope microfracture as center, construct spatial heterogeneity mapping;Utilize convolutional neural network to extract vegetation root system competition characteristics, determine multi-species nutrient absorption demand mode;The above information is integrated into three-dimensional voxel model, and combined with hydrology model, simulate the migration path of water and fertilizer along the fracture, and predict the risk of nutrient enrichment downhill in advance;Finally generate a three-dimensional voxel model adjustment scheme for slope, soil moisture content and predicted loss path for fine adjustment, to guide precision variable fertilization and fertility repair.The application also provides a kind of device for executing the method.The application can significantly improve the precision, forward-looking and timeliness of precision fertilization and fertility management under complex terrain.

Description

Technical Field

[0001] This invention relates to the field of agricultural soil remediation technology, and more specifically, to a method and apparatus for precise variable fertilization of farmland soil that integrates multimodal data. Background Technology

[0002] Soil fertility is a core material foundation for ensuring food security and sustainable agricultural development. Existing precision agriculture technologies, such as Variable Rate Application (VRA), have achieved some success in improving fertility management in plains areas and monoculture planting patterns. However, these technologies typically rely on two-dimensional planar representations of the soil environment, such as remote sensing imagery based on the Normalized Difference Vegetation Index (NDVI) or gridded soil sampling data, which are insufficient to effectively address the challenges of fertility management in areas with complex terrain features.

[0003] When application scenarios shift from plains to mountains and hills, the spatial distribution of soil fertility becomes extremely complex. Under the combined influence of factors such as gravity, rainfall erosion, and slope aspect, soil nutrients not only exhibit high heterogeneity in the horizontal direction but also form complex dynamic stratification and migration in the vertical depth. Current technologies lack effective means of sensing, modeling, and regulating this, making it impossible to truly achieve spatiotemporal dynamic fertility restoration. This often leads to water and fertilizer migration and loss along the slope, exacerbating soil degradation and posing a potential threat to the downstream ecological environment.

[0004] This problem is particularly prominent in intercropping systems of high-value cash crops and soil-stabilizing green manure plants in mountainous areas. In such complex systems, soil fertility is the result of the coupled influence of multiple factors, including topography, climate conditions, and the interaction of root systems of various species, exhibiting highly dynamic spatiotemporal characteristics. Data from a single source is no longer sufficient to accurately characterize soil conditions. For example, remote sensing imagery alone is insufficient to determine the distribution of underground roots and the status of deep nutrients; while ground sensor data alone cannot provide overall macroscopic spatial distribution characteristics. In particular, topographic features such as micro-fissures on slopes, as advantageous channels for rapid water and fertilizer loss, often lead to degradation problems caused by excessive local nutrient loss.

[0005] Therefore, existing technologies struggle to generate high-resolution three-dimensional variable fertilizer prescription maps, which limits the accuracy of fertilization operations. Summary of the Invention

[0006] The technical problem this invention aims to solve is to overcome the shortcomings of existing technologies in handling complex three-dimensional systems such as mountain intercropping, which result in low soil fertility management accuracy and the inability to achieve precise variable fertilization based on risk prediction due to insufficient model representation dimensions and inadequate data fusion. To this end, this invention provides a method and apparatus for precise variable fertilization of farmland soil by integrating multimodal data.

[0007] To achieve the above objectives, one aspect of the present invention provides a method for precise variable fertilization of farmland soil by integrating multimodal data, the method comprising the following steps:

[0008] S1. Collect the data required for spatial heterogeneity mapping and perform preliminary mapping. This step involves deploying a soil sensor network around the micro-fractures on the slope of the mountain intercropping system according to preset rules to obtain high spatiotemporal resolution data on soil moisture and nutrient distribution. This data is then combined with topographic parameters such as slope and aspect obtained from a Digital Elevation Model (DEM) to form a preliminary spatial dataset. Subsequently, the spatial analysis function of a Geographic Information System (GIS) is used to perform raster overlay processing on the slope topographic parameters and soil moisture and nutrient distribution data, delineating the fracture-affected areas. The coefficient of variation of nutrient concentration within these areas is statistically calculated to obtain quantified heterogeneity indicators, ultimately generating a preliminary mapping reflecting the spatial heterogeneity under the influence of micro-fractures on the slope.

[0009] This invention, by deploying sensors centered on micro-cracks in the slope and collecting data, can obtain an initial spatial heterogeneity mapping characterizing key pathways of water and fertilizer loss, providing a data foundation with clear physical meaning for subsequent analysis.

[0010] S2, Extracting Root Competition Features and Determining Nutrient Absorption Demand Patterns. This step, based on the preliminary spatial heterogeneity mapping generated in step S1, integrates vegetation root depth scan data obtained using ground-penetrating radar or root drilling methods to form a root distribution dataset containing spatial location information. A Convolutional Neural Network (CNN) is used to process this dataset. Its convolutional layers automatically learn and capture local patterns such as the interlacing and overlapping of different species' root systems in three-dimensional space. After processing through pooling and fully connected layers, quantified root competition features are extracted. These features are then fused with soil type distribution data obtained from a pre-defined database. When the root competition feature value exceeds a pre-defined threshold, the soil type data weights are adjusted to calculate the multi-species competition intensity that comprehensively considers the rhizosphere environment, and based on this, the nutrient absorption demand patterns of multiple species are determined.

[0011] S3, Integrating Multidimensional Data to Generate a Moisture Attribute Distribution Map. This step integrates the multi-species nutrient absorption demand patterns determined in step S2 with a voxel model representing the three-dimensional structure of the soil, and combines this with soil physicochemical properties such as soil texture and organic matter content to obtain preliminary attribute integration data. Further, soil nutrient content attributes are introduced, and their interaction strength with the preliminary integrated data is calculated. When the interaction strength exceeds a preset threshold, the weights of the nutrient content attributes are adjusted to generate enhanced interaction results. Using these enhanced interaction results, and specifically for micro-fracture areas on slopes, another convolutional neural network model is used to simulate and predict preferred water infiltration paths. Finally, combined with spatial heterogeneity mapping, a three-dimensional moisture attribute distribution map that can accurately characterize the water movement trend under the influence of fractures is generated.

[0012] By using convolutional neural networks to extract root competition features from root scan data and integrating them with soil physicochemical properties into a three-dimensional voxel model, the model can simulate nutrient dynamics caused by interspecies interactions.

[0013] S4 involves simulating water and nutrient migration and predicting the risk of nutrient enrichment on the downhill slope. This step first analyzes the moisture distribution map generated in step S3 to determine if the root density in the uphill area is below a preset soil-fixing capacity threshold. If it is below this threshold, it indicates a high risk of soil erosion. In this case, real-time climate monitoring data such as rainfall and intensity from meteorological stations are further integrated, and combined with a hydrological model based on physical processes. Numerical methods such as the finite difference method are used to solve the problem, generating a refined water and nutrient migration simulation path. Finally, combined with soil moisture properties, the risk of excessive nutrient enrichment in the downstream slope area is assessed and quantified.

[0014] S5, Optimize fertilization distribution parameters and generate a 3D voxel model adjustment scheme. This step, based on the water and fertilizer migration simulation path generated in step S4, uses a soil physics model to solve for water movement in unsaturated soil, capturing soil infiltration characteristics under different slopes. Combined with the crop's fertilizer absorption rate attribute, a weighted calculation is performed to optimize and generate differentiated fertilization distribution parameters. These parameters are mapped onto a 3D voxel grid to construct a 3D voxel model adjustment scheme that finely adjusts the slope, soil moisture content, and predicted loss paths. This scheme clearly defines the optimization suggestions for fertilizer type, quantity, and location in 3D space.

[0015] Preferably, after generating the three-dimensional voxel model adjustment scheme, the method of the present invention further includes a high-precision fertilization execution step. Specifically, firstly, according to the three-dimensional voxel model adjustment scheme, standardized fertilization units are prepared by mixing solid slow-release fertilizer, water-retaining agent, or soil conditioner in different proportions. Then, using a high-precision pneumatic or piezoelectric soil injection device linked to the three-dimensional voxel model and equipped with a high-precision positioning system, a specific number and proportion of fertilization units are precisely delivered to the soil spatial coordinates (X, Y, Z) positions specified by the three-dimensional voxel model. This injection process can deploy fertilization units at different soil depths, especially at key nodes of the water and fertilizer migration path predicted by the model or in areas with weak root systems, thereby achieving high-precision three-dimensional spatial point-to-point fertilization.

[0016] After generating the three-dimensional voxel model adjustment scheme, the method of the present invention may further include a zoned differentiated comprehensive land remediation step. Specifically, based on the three-dimensional voxel model adjustment scheme and the water and fertilizer migration simulation path, clustering algorithms are used to automatically zon the farmland surface, dividing it into at least three types of areas: uphill nutrient loss risk zone, mid-slope stable zone, and downhill nutrient over-accumulation risk zone. For different zones, automated agricultural machinery linked to the zoning map performs differentiated land remediation measures. For example, in the uphill nutrient loss risk zone, surface covering with organic matter (such as straw) is used in conjunction with spraying biodegradable soil stabilizing agents to slow down water flow and nutrient loss; in the downhill nutrient over-accumulation risk zone, fertilization is reduced or stopped, and specific high-nutrient-absorbing plants can be planted to consume excess nutrients.

[0017] S6, Determine the nutrient allocation equilibrium point under the influence of root competition. This step integrates historical crop growth records and the root competition influence determined in step S2 by adjusting the three-dimensional voxel model to generate preliminary nutrient allocation parameters. Based on these parameters, the dynamic distribution of soil nutrients is simulated, and a competition assessment mechanism value that quantifies the degree of competition is extracted. If this value is greater than a preset threshold, a competition influence matrix is ​​constructed through data fusion and iterative optimization, ultimately determining a theoretical nutrient allocation equilibrium point that enables the coordinated growth of multiple species.

[0018] S7, Verify and Generate the Final Optimized Scheme. This step employs a three-dimensional convolutional neural network (3D CNN), using the nutrient allocation equilibrium point determined in step S6 as one of the input conditions to simulate the future dynamics of the soil system, thereby verifying the adaptability of this equilibrium point to spatial heterogeneity and changes in root density. By analyzing the quantitative indicators of competition effects in the simulation results, the allocation scheme is fine-tuned, ultimately generating a final optimized scheme that balances long-term benefits and system stability. This scheme can be used to guide agricultural production activities in the next growth cycle.

[0019] In another aspect, the present invention provides a precision variable fertilization device for farmland that integrates multimodal data. This device is used to perform the above-described method and may include:

[0020] The data acquisition module is used to perform step S1, which involves collecting multimodal data such as soil and topography through sensor networks and other data sources.

[0021] The model building and analysis module is equipped with a convolutional neural network, a hydrological model, and a soil physics model. It is used to execute steps S2 to S5 to realize spatial heterogeneity mapping, root competition feature extraction, water attribute distribution map generation, water and fertilizer migration simulation, and generation of three-dimensional voxel model adjustment scheme.

[0022] The precision fertilization execution module may optionally include a high-precision soil injection device or an automated agricultural machinery control unit linked to the model building and analysis module, for performing the high-precision physical remediation or zoned differentiated treatment.

[0023] Compared with existing technologies, this invention, firstly, fundamentally overcomes the inherent limitations of traditional two-dimensional characterization methods in dealing with complex terrains such as mountains and hills by constructing a three-dimensional voxel model that integrates topography, spatial distribution of root systems of multiple species, and soil physicochemical properties. This three-dimensional model not only achieves high-precision characterization of the spatial heterogeneity of the soil environment, but also innovatively employs techniques such as convolutional neural networks to transform the originally difficult-to-quantify inter-root competition relationships into calculable model parameters. This allows it to realistically reflect the dynamic interactions of nutrients in multi-species symbiotic systems, significantly improving the comprehensiveness and accuracy of soil fertility status assessment.

[0024] Secondly, this invention achieves forward-looking risk prediction and precise intervention. By combining hydrological and soil physics models, this invention can simulate water and fertilizer migration paths under the influence of external factors such as rainfall, thus predicting the risk of excessive enrichment in downstream areas before nutrient loss occurs. This predictive capability transforms fertility management from passive, delayed remediation to proactive, preventative, closed-loop regulation. Based on this prediction, the generated three-dimensional voxel model adjustment scheme can pinpoint remediation instructions to specific coordinates and depths in three-dimensional space, guiding the precise delivery of fertilizer units, thereby maximizing nutrient utilization efficiency while minimizing negative environmental impacts.

[0025] Finally, by incorporating historical data and simulating, determining, and verifying the nutrient distribution equilibrium point, the final optimized scheme generated by this invention is no longer a static, one-time fertilization guide, but a dynamic strategy that takes into account the synergistic growth of multiple species, long-term economic benefits, and system stability. This scheme has been validated for adaptability to different climatic conditions (such as drought and abundant rainfall), exhibiting stronger robustness and providing agricultural producers with a scientific basis for decision-making. Attached Figure Description

[0026] Figure 1 This is a flowchart of a precise variable fertilization method for farmland soil that integrates multimodal data, provided by an embodiment of the present invention.

[0027] Figure 2 This is a schematic diagram illustrating the principle of generating a three-dimensional moisture attribute distribution map by fusing multi-dimensional data in an embodiment of the present invention.

[0028] Figure 3 This is a structural block diagram of a precision variable fertilization device for farmland that integrates multimodal data, provided in an embodiment of the present invention.

[0029] Figure 4 This is a schematic diagram of three-dimensional fusion of multimodal data of the mountain intercropping system in an embodiment of the present invention.

[0030] Figure 5 This is a schematic diagram of water and fertilizer migration path simulation and risk prediction based on a physical model in an embodiment of the present invention.

[0031] Figure 6 This is a schematic diagram of the three-dimensional voxelization precision repair scheme in an embodiment of the present invention. Detailed Implementation

[0032] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention.

[0033] Example 1

[0034] This invention provides a method for precise variable fertilization of farmland soil by fusing multimodal data. This method addresses the problems of low fertility management accuracy and the inability to achieve proactive remediation in complex three-dimensional systems such as mountainous terrain due to insufficient model dimensions and data fusion in existing technologies. (Refer to...) Figure 1 The method may include the following steps:

[0035] Step S1: Collect the data required for spatial heterogeneity mapping and perform preliminary mapping.

[0036] This step involves deploying a soil sensor network around micro-fractures on the slope of the mountain intercropping system according to pre-defined rules to acquire high spatiotemporal resolution data on soil moisture and nutrient distribution. This data is then combined with topographic parameters such as slope and aspect obtained from a Digital Elevation Model (DEM) to form a preliminary spatial dataset. Subsequently, the spatial analysis function of a Geographic Information System (GIS) is used to rasterize and overlay the slope topographic parameters and soil moisture and nutrient distribution data, delineating the fracture-affected areas. The coefficient of variation of nutrient concentration within these areas is then statistically calculated to obtain a quantified heterogeneity index, ultimately generating a preliminary mapping reflecting the spatial heterogeneity under the influence of micro-fractures on the slope.

[0037] Specifically, the sensor network is not deployed uniformly or randomly, but rather with a gradient density deployment strategy centered on micro-cracks in the slope. For example, within a range of 0-0.5 meters from the crack edge, the sensor deployment density can be set to one sensor per square meter; within a range of 0.5-2 meters, the density can be reduced to one sensor per 4 square meters. The soil sensors used may include time-domain reflectometry (TDR) sensors for measuring volumetric water content, and ion-selective electrode sensors for measuring the concentrations of nitrogen (N), phosphorus (P), and potassium (K) ions in the soil solution. The data acquisition frequency can be set to once per hour to capture dynamic changes caused by events such as rainfall. The digital elevation model (DEM) data can be obtained from UAV LiDAR (Light Detection and Ranging) scans, with a spatial resolution preferably of 0.5 m × 0.5 m or higher.

[0038] In GIS processing, commercial software such as ArcGIS or open-source software QGIS can be used. First, point data collected by sensors is interpolated using Kriging to generate continuous soil moisture and nutrient distribution layers. Then, slope and aspect layers are extracted using DEM data. Overlay analysis tools are used to correlate areas with slopes greater than a specific threshold (e.g., 15 degrees) and nutrient concentrations lower than crop requirement thresholds (e.g., nitrate nitrogen below 10 mg / kg) with fissure locations, identifying high-risk areas for nutrient loss due to fissure influence. The heterogeneity index, namely the coefficient of variation (CV), is calculated using the formula CV = (σ / μ) × 100%, where σ is the standard deviation of nutrient concentration within the region, and μ is the average. When the CV value exceeds a preset threshold, such as 30%, the region is marked as highly heterogeneous. The principle behind this step is to prioritize the identification and quantification of nutrient unevenness caused by micro-fractures, a key physical structure. This provides subsequent models with high-value initial boundary conditions that have clear physical meaning, enabling the identification and quantification of key drivers of water and fertilizer loss from a physical perspective. Compared to traditional macroscopic sampling, this improves the targeting and accuracy of problem diagnosis. As an alternative, during the data acquisition phase, drones equipped with multispectral or hyperspectral cameras can be used to analyze the correlation between vegetation indices (such as NDVI and NDRE) and data from a small number of calibration points in the field, thereby retrieving the nutrient distribution of the topsoil over a large area and reducing sensor deployment costs.

[0039] Step S2: Extract root competition characteristics and determine nutrient absorption demand patterns.

[0040] This step, based on the preliminary spatial heterogeneity mapping generated in step S1, integrates vegetation root depth scan data obtained using Ground Penetrating Radar (GPR) or root drilling methods to form a root distribution dataset containing spatial location information. A Convolutional Neural Network (CNN) is used to process this dataset. Its convolutional layers automatically learn and capture local patterns such as the interlacing and overlapping of root systems of different species in three-dimensional space. After further processing by pooling and fully connected layers, quantified root competition features are extracted. These features are then fused with soil type distribution data obtained from a pre-defined database. When the root competition feature value exceeds a preset threshold, the soil type data weights are adjusted to calculate the multi-species competition intensity that comprehensively considers the rhizosphere environment, and based on this, the nutrient absorption demand patterns of multiple species are determined.

[0041] In a specific application scenario, such as on a slope where citrus and clover are intercropped, a GPR device is used to scan the target area with a 0.2m × 0.2m grid to obtain radar echo maps within a depth range of 0-1.5m. These maps are processed into 3D data cubes and used as input to a CNN model. This CNN model can employ an encoder-decoder structure similar to U-Net, where the encoder extracts the morphological and spatial distribution features of the root system layer by layer through successive 3D convolutional layers (e.g., using 3×3×3 convolutional kernels) and max pooling layers; the decoder, through upsampling and convolution operations, ultimately outputs a 3D segmentation map of the same size as the input, marking the location and density of the root systems of different species. From this segmentation map, the proportion of root volumes of different species within a specific voxel can be calculated, thus obtaining quantified root competition feature values. For example, a competition index RCI = V_A / (V_A + V_B) can be defined, where V_A and V_B are the root volumes of the two crops, respectively. When the RCI deviates from 0.5 by more than a threshold (e.g., greater than 0.7 or less than 0.3), it indicates the presence of a competitive relationship. In this case, the feature is fused with soil type data (e.g., red soil, sandy loam) obtained from the *China Soil Database*. If strong root competition is detected in sandy loam areas, the weight of the soil attribute with weak water and fertilizer retention capacity is increased when calculating nutrient requirements, thus more accurately estimating actual nutrient availability. The technical effect of this step is that it transforms the invisible underground biological process (root competition) into a calculable quantitative indicator and deeply couples it with the soil physical environment, making nutrient requirement assessment no longer based on an idealized model of a single crop, but reflecting the real dynamics of a multi-species symbiotic system. As an alternative, besides CNNs, traditional image processing algorithms based on texture analysis (such as Gabor filters) can be used to extract root features, or machine learning models such as random forests can be used to classify and quantify root distribution.

[0042] Step S3: Integrate multidimensional data to generate a moisture attribute distribution map.

[0043] This step integrates the multi-species nutrient uptake demand patterns determined in step S2 with a voxel model characterizing the three-dimensional structure of the soil, and combines this with soil physicochemical properties such as soil texture and organic matter content to obtain preliminary attribute integration data. Further, soil nutrient content attributes are introduced, and their interaction strength with the preliminary integration data is calculated. When the interaction strength exceeds a preset threshold, the weights of the nutrient content attributes are adjusted to generate enhanced interaction results. Using these enhanced interaction results, and targeting micro-fracture areas on slopes, another convolutional neural network model is used to simulate and predict preferential water infiltration paths. Finally, combined with spatial heterogeneity mapping, a three-dimensional water attribute distribution map that can accurately characterize the water movement trend under the influence of fractures is generated.

[0044] Reference Figure 2 and Figure 4 The first step involves constructing a three-dimensional voxel model covering the entire study area. For example, the voxel resolution can be set to 0.25 m × 0.25 m × 0.1 m. Each voxel is assigned multiple attributes, including topographic parameters (slope) obtained in step S1, root density and competition index obtained in step S2, and soil texture (percentage of sand, silt, and clay) and organic matter content obtained from soil databases or measured data. These attributes constitute a multi-channel input tensor. Subsequently, measured soil nutrient content (N, P, K concentration) is introduced as another attribute. The interaction strength can be calculated as a weighted summation process, for example, interaction strength I = w1*f(root competition) + w2*f(organic matter) + ..., where w is the weight and f is the normalization function. When the I value exceeds a preset threshold (e.g., 0.8), it indicates that multiple factors within the voxel have a strong combined influence on nutrient availability, and the weight of the nutrient content attribute of that voxel in subsequent calculations is increased.

[0045] The CNN model takes voxel properties, incorporating the enhanced interaction results mentioned above, as input and outputs a three-dimensional vector field, where each vector represents the direction and relative rate of water flow within the corresponding voxel. Trained on a large amount of data based on physical simulations (e.g., generated using HYDRUS-3D software) or experimental tracer studies, the model captures how fractures alter soil hydraulic properties. Finally, this predicted preferred infiltration path (vector field) is fused with the spatial heterogeneity map generated in step S1 to generate the final three-dimensional water property distribution map. This map not only demonstrates the static water distribution potential but, more importantly, dynamically reveals how water and dissolved nutrients preferentially migrate through the fracture network during events such as rainfall. The technical advantage of this step lies in replacing or enhancing traditional, computationally expensive physical simulations with a data-driven CNN model, achieving rapid and accurate prediction of preferred flows in a heterogeneous body like fractures. As an alternative, instead of using CNN, a porous medium flow model based on the finite element method or finite difference method (such as the Richards equation solver) can be used for simulation. However, the computational complexity and the requirements for parameter accuracy will be significantly increased.

[0046] Step S4: Conduct water and fertilizer migration simulation and predict the risk of enrichment on the downhill slope.

[0047] This step first analyzes the moisture distribution map generated in step S3 to determine whether the root density in the upslope area is below a preset soil-fixing capacity threshold. If it is below this threshold, it indicates a high risk of soil erosion in the area. In this case, real-time climate monitoring data such as rainfall and rainfall intensity provided by the meteorological station are further integrated, and combined with a hydrological model based on physical processes, numerical methods such as the finite difference method are used to solve the problem and generate a refined simulation path for water and fertilizer migration. Finally, combined with soil moisture content attributes, the risk of nutrient over-accumulation in the downstream slope area is determined and quantified.

[0048] In practice, the soil stabilization capacity threshold can be set according to the specific vegetation type. For example, for clover with shallow root systems, if the average root density in the 0-20 cm soil layer is less than 5 kg / m³, it can be determined that the soil stabilization capacity is insufficient. When a certain area on the uphill slope is identified as a risk zone, the system will automatically trigger the migration simulation module. This module connects to a local small weather station or a national meteorological data interface to obtain the rainfall forecast for the next 24 hours, such as if the forecast rainfall intensity exceeds 10 mm / hour. At this time, a slope flow model based on the 2D Saint-Venant equations is started. The input of this model includes the topography provided by the DEM, the soil infiltration rate predicted in step S3 (derived from the water property distribution map), and rainfall data. By solving the equations in time and space using the finite difference method, the velocity, direction, and depth of surface runoff can be simulated. At the same time, by coupling the concentration of dissolved nutrients with runoff and solving the convection-dispersion equations, a simulated path map of water and fertilizer migration can be generated. This path map clearly shows nutrient loss from the uphill slope and potential accumulation at topographical changes such as the foot of the slope and the edges of terraces. By calculating the predicted total nutrient input to the simulated path's endpoint area and comparing it with the current environmental capacity of the soil in that area (determined by soil type and existing nutrient content), the risk of nutrient over-accumulation can be quantified. For example, if the predicted nitrogen input exceeds 20% of the downstream soil's retention capacity, it is considered a high risk of over-accumulation. This step represents a shift from static diagnosis to dynamic prediction; it not only identifies where nutrients are deficient but also anticipates where nutrients will flow after fertilizer application or natural rainfall, thus providing a basis for proactive intervention rather than passive remediation. The simulated water and fertilizer migration paths and risk prediction results can be found in [reference needed]. Figure 5 As shown.

[0049] Step S5: Optimize fertilizer distribution parameters and generate a three-dimensional voxel model adjustment plan.

[0050] This step, based on the simulated water and fertilizer migration path generated in step S4, uses a soil physics model to solve for water movement in unsaturated soil, capturing soil infiltration characteristics under different slopes. It also incorporates crop fertilizer absorption rate attributes for weighted calculations, thereby optimizing and generating differentiated fertilization distribution parameters. These parameters are mapped onto a three-dimensional voxel mesh to construct a three-dimensional voxel model adjustment scheme that finely adjusts the slope, soil moisture content, and predicted loss paths. This scheme clearly defines optimization suggestions for fertilizer type, quantity, and location in three-dimensional space.

[0051] The core of this step is solving the soil moisture movement model based on the Richards equations: Where h is the head, t is time, C(h) is the specific water capacity, and K(h) is the hydraulic transfer function. This is the gradient operator. These parameters (C(h) and K(h)) can be determined using empirical models such as van Genuchten-Mualem, based on soil texture, organic matter content, and other attributes of each voxel in step S3. By solving this equation, the vertical and horizontal movement of water and nutrients within each voxel after fertilization can be accurately simulated. Combining the crop root distribution and nutrient uptake rate determined in step S2 (which can be obtained from the agricultural knowledge base), the system performs an optimization calculation. The objective function of this optimization can be set as: maximizing nutrient availability within the root zone while minimizing nutrient flux along the predicted loss path. The optimization algorithm can employ a genetic algorithm or a particle swarm optimization algorithm. Its output is a set of differentiated fertilization distribution parameters. For example, for voxels in the uphill loss risk zone, it is recommended to apply controlled-release nitrogen fertilizer at a depth of 15 cm; for the stable mid-slope zone, fast-acting compound fertilizer can be used at a depth of 10 cm; for the downhill enrichment risk zone, it is recommended not to fertilize or to apply micronutrient fertilizer. These parameters are precisely mapped back to the three-dimensional voxel model, forming an intuitive and executable three-dimensional fertilization prescription diagram, i.e., a three-dimensional voxel model adjustment scheme (its schematic diagram can be found in [reference]). Figure 6 This solution enables precise control in three-dimensional space (X, Y, Z), which can fundamentally solve the problem of water and fertilizer loss along the dominant path after fertilization on slopes.

[0052] Preferably, after generating the three-dimensional voxel model adjustment scheme, the method of the present invention further includes a high-precision fertilization execution step. Specifically, firstly, according to the three-dimensional voxel model adjustment scheme, solid slow-release fertilizers (such as coated urea), water-retaining agents (such as potassium polyacrylate), or soil conditioners (such as biochar) with different ratios are made into standardized spherical or cylindrical fertilization units with a diameter of 1-2 cm. Then, using a pneumatic or piezoelectric soil injection device linked to the three-dimensional voxel model and equipped with a real-time kinematic global positioning system (RTK-GPS) high-precision positioning system (positioning accuracy up to the centimeter level), a specific number and ratio of fertilization units are precisely delivered to the soil spatial coordinates (X, Y, Z) position specified by the three-dimensional voxel model. For example, the device can be installed on a small tracked robot or extended from the field ridge by a cantilever. Its injection probe can accurately deploy the fertilization unit at different depths such as 5 cm, 15 cm or 30 cm underground according to the program instructions, especially at key nodes of the water and fertilizer migration path predicted by the model or in areas with weak root systems, thereby achieving precise point repair to the predetermined three-dimensional coordinates.

[0053] Furthermore, after generating the three-dimensional voxel model adjustment scheme, the method of the present invention may also include a zone-differentiated comprehensive surface management step. Based on the three-dimensional voxel model adjustment scheme and the water and fertilizer migration simulation path, clustering algorithms such as K-Means or DBSCAN are used to automatically zone the farmland surface, dividing it into at least three types of areas: uphill nutrient loss risk zone, mid-slope stable zone, and downhill nutrient over-accumulation risk zone. For different zones, automated agricultural machinery linked to the zoning map performs differentiated surface remediation measures. For example, in the uphill nutrient loss risk zone, a variable displacement mulch machine is instructed to lay a layer of straw or sawdust 5-10 cm thick on the surface, and a drone is used to spray biodegradable polyacrylamide (PAM) soil stabilizing agent to increase surface roughness and slow down runoff velocity and nutrient loss; in the downhill nutrient accumulation risk zone, a variable displacement fertilizer applicator is instructed to stop fertilization in that area, and a seeder can replant highly nutrient-absorbing plants such as ryegrass in that area to consume and transfer any excess nutrients that may have accumulated in the soil.

[0054] Step S6: Determine the nutrient distribution equilibrium point under the influence of root competition.

[0055] This step integrates historical crop growth records (such as annual yields and canopy size) through the adjustment scheme of the three-dimensional voxel model, and combines them with the root competition influence determined in step S2 to generate preliminary nutrient allocation parameters. Based on these parameters, the dynamic distribution of soil nutrients is simulated, and a competition assessment mechanism value that quantifies the degree of competition is extracted. If this value is greater than a preset threshold, a competition influence matrix is ​​constructed through data fusion and iterative optimization, ultimately determining a theoretically balanced nutrient allocation point that enables the coordinated growth of multiple species.

[0056] Specifically, the initial nutrient allocation parameters are based on fertilization recommendations adjusted from the 3D voxel model, taking into account the different crop responses to nutrients reflected in historical data. For example, historical data shows that citrus yields increase significantly when phosphorus fertilizer is sufficient, while clover is more sensitive to nitrogen fertilizer. The system performs initial allocation based on this and simulates nutrient consumption over a growth cycle. During this process, the root competition index (RCI) defined in step S2 is continuously calculated as a competition evaluation mechanism value. If the RCI of a certain region remains greater than 0.8 during the simulation (indicating that citrus roots excessively encroach on clover space), iterative optimization is initiated. The optimization process constructs a competition influence matrix M, where the element M_ij represents the inhibition coefficient of species j on species i in the absorption of a specific nutrient (such as nitrogen). This matrix is ​​gradually corrected by comparing the simulated growth results under different nutrient allocation schemes. Finally, a theoretical nutrient allocation equilibrium point is determined by solving a multi-objective optimization problem (e.g., maximizing total economic benefits while constraining the competition indices of both crops to remain within the synergistic growth range of 0.4-0.6). This equilibrium point may manifest as a more refined scheme of N, P, K ratios and application timing.

[0057] Step S7: Verify and generate the final optimized solution.

[0058] This step employs a three-dimensional convolutional neural network (3D CNN), using the nutrient allocation equilibrium point determined in step S6 as one of the input conditions to simulate the future dynamics of the soil system, thereby verifying the adaptability of this equilibrium point to spatial heterogeneity and changes in root density. By analyzing the quantitative indicators of competition effects in the simulation results, the allocation scheme is fine-tuned, ultimately generating a final optimized scheme that balances long-term benefits and system stability. This scheme can be used to guide agricultural production activities in the next growth cycle.

[0059] The 3D CNN here can be viewed as a digital twin surrogate model trained to predict the three-dimensional nutrient distribution, water status, and crop root growth of the soil after one growing cycle (e.g., 3 months) given initial soil conditions, climate conditions, and management practices (i.e., nutrient allocation schemes). The model's inputs include: the current three-dimensional voxel model, the nutrient allocation equilibrium point scheme determined in step S6, and a probabilistic prediction of future climate (e.g., drought-prone, normal, or waterlogged). After running, the model outputs a predicted future three-dimensional soil state. This prediction result is analyzed to check whether indicators such as the root competition index, nutrient utilization rate, and predicted loss are still within ideal ranges. If the prediction shows that in drought-prone years, the equilibrium point scheme will lead to increased root competition in a certain area, the system will fine-tune the scheme; for example, by appropriately increasing the amount of deep water-retaining agent in that area during drought warnings. By simulating extreme climatic conditions (such as drought or waterlogging) to validate the scheme, the final optimized scheme improved its robustness and can guide farmers to conduct forward-looking and adaptive long-term fertility management.

[0060] Example 2

[0061] This invention also provides a precision variable fertilization device for farmland that integrates multimodal data, used to perform the method described in Embodiment 1. (Refer to...) Figure 3 The device 100 can be a server, an embedded system, or a cloud computing platform.

[0062] The device 100 may include at least one processor 101 and a memory 102 communicatively connected to the at least one processor 101.

[0063] The memory 102 stores computer program instructions, which, when executed by the at least one processor 101, cause the device 100 to perform the method described in Embodiment 1.

[0064] Preferably, the functional units of the device 100 can be divided into:

[0065] The data acquisition module 103 is used to perform step S1. It may include a data interface for connecting to external data sources such as soil sensor networks, drones, and weather stations, and preprocess, format, and store the received multimodal data such as soil, topography, and weather.

[0066] The model building and analysis module 104 is implemented by the processor 101 executing specific program instructions from the memory 102. Internally, it includes a convolutional neural network (CNN) for extracting root features, a hydrological model for simulating water and fertilizer migration (such as a model based on the Saint-Venant equations), and a soil physics model for optimizing fertilization parameters (such as a model based on the Richards equations). This module is responsible for performing the core computational tasks of steps S2 to S5, realizing spatial heterogeneity mapping, root competition feature extraction, water attribute distribution map generation, water and fertilizer migration simulation, and the generation of a three-dimensional voxel model adjustment scheme. This module may require high-performance computing resources, such as graphics processing units (GPUs), to accelerate the training and inference of the neural network.

[0067] The scheme optimization and verification module 105 is responsible for executing steps S6 and S7. It receives the three-dimensional voxel model adjustment scheme and integrates historical crop growth record data with the root system competition influence to determine the nutrient distribution equilibrium point for multi-species synergistic growth. Furthermore, the module uses a three-dimensional convolutional neural network to simulate and verify the long-term adaptability of the equilibrium point, and finally generates the final optimization scheme.

[0068] The precision fertilization execution module 106 is an optional module configured when the device needs to directly control physical equipment. As an output control unit, it connects with the control system of high-precision soil injection equipment or automated agricultural machinery (such as variable displacement fertilizer applicators or drones) via wired or wireless communication (e.g., LoRa, 4G / 5G). This module is responsible for converting the 3D voxel model adjustment scheme or zoning treatment map generated by the model building and analysis module 104 into execution instructions for specific equipment (such as coordinates, injection depth, fertilizer application rate, spraying rate, etc.), thereby achieving closed-loop automated precision fertilization operations.

[0069] Those skilled in the art will understand that the accompanying drawings are merely schematic diagrams of a preferred embodiment, and the modules or processes shown in the drawings are not necessarily essential for implementing the present invention.

[0070] Those skilled in the art will understand that the modules in the apparatus of the embodiments may be distributed in the apparatus of the embodiments as described in the embodiments, or may be located in one or more devices different from this embodiment with corresponding changes. The modules of the foregoing embodiments may be combined into one module, or may be further divided into multiple sub-modules.

[0071] Example 3

[0072] Taking a terraced tea garden located in subtropical hilly area A as an example, the tea garden intercrops green manure crop B between the tea tree rows to improve the soil. The soil type in this area is typical acidic red soil, which is susceptible to water erosion during the rainy season, especially forming micro-cracks on the terrace walls, leading to water and fertilizer loss.

[0073] In this scenario, the spatial heterogeneity mapping data was first collected. A T800 UAV equipped with a LiDAR scanner scanned the entire tea garden, generating a digital elevation model (DEM) with a resolution of 0.1 m × 0.1 m. Analysis of the DEM accurately identified micro-cracks exceeding 0.5 m in length on the terrace walls. Subsequently, an S-shaped soil sensor array was deployed around these cracks using a gradient density approach: one sensor per 0.25 square meters within a 0.2 m radius of the crack, decreasing to one sensor per 1 square meter between 0.2 and 1 m. The sensors collected real-time data on soil nitrate nitrogen, available phosphorus, available potassium content, and volumetric water content. The collected data underwent Kriging interpolation and overlay analysis using Geographic Information System (GIS) software. The results showed that after rain, the coefficient of variation (CV) of nitrate nitrogen concentration within a 0.5 m radius of the cracks reached as high as 45%, quantifying the highly uneven nutrient distribution caused by the cracks.

[0074] Next, to extract root competition features, a GPR-3000 ground-penetrating radar was used to scan the tea garden soil with a 0.3m × 0.3m grid at a depth of 1.2m. The acquired radar echo images were processed into three-dimensional data cubes and input into a convolutional neural network (CNN) model based on the U-Net architecture. This pre-trained model was able to accurately segment the taproot system of the tea trees and the fibrous root system of the green manure crop B, and calculate the root volume percentage within each 0.1m × 0.1m × 0.1m voxel, thus obtaining a quantified root competition index (RCI). In areas with dense tea tree roots, the RCI values ​​were generally higher than 0.75, indicating that the tea trees dominate nutrient absorption in this area.

[0075] Subsequently, the system integrates multi-dimensional data to generate a moisture attribute distribution map. A three-dimensional voxel model covering the entire tea garden was established, with a voxel resolution of 0.2 m × 0.2 m × 0.1 m. Each voxel was assigned multiple attributes, including slope, root competition index, soil type (red soil), organic matter content, and measured nutrient concentration. A CNN model trained on physically simulated data was used to process the voxel data containing the above attributes; this model is specifically designed to predict preferred infiltration pathways. The three-dimensional moisture attribute distribution map output by the model clearly shows that, under simulated rainfall intensity of 30 mm / h, moisture and dissolved nutrients will preferentially migrate down to the lower layers and adjacent terraces along several major fissure pathways.

[0076] Based on this, the system simulates and predicts water and fertilizer migration risks. According to the water distribution map and root density data, some uphill areas are identified as having a high risk of nutrient loss due to insufficient soil-fixing capacity of green manure roots (root density below 4 kg / m³). Combined with heavy rain warnings provided by the meteorological station, a hydrological model based on the two-dimensional Saint-Venant equations is activated. Simulation results show that without intervention, in a typical heavy rain event, approximately 22% of the nitrogen fertilizer applied in the uphill area will be lost through a specific dominant fissure, causing nutrient over-accumulation near the ridges of the downhill terraces, with the predicted accumulation exceeding 30% of the soil's environmental capacity.

[0077] Ultimately, the system generates a three-dimensional voxel model adjustment plan to guide precise restoration. Using a model based on the Richards equations, with the optimization objectives of maximizing nutrient availability in the tea tree root zone and minimizing nutrient loss through fissure pathways, differentiated fertilization parameters are calculated. The plan specifies that for voxels in identified high-risk loss areas, 40 grams of a fertilization unit composed of a mixture of biochar, humic acid, and coated urea should be injected at a depth of 25 cm; for nutrient-rich risk areas on downhill slopes, nitrogen and phosphorus fertilizer application should be stopped, and potassium and silicon fertilizers should be increased to enhance the tea tree's resistance. This plan is then sent to a field operation robot equipped with a real-time dynamic differential positioning (RTK) system. Based on three-dimensional coordinate instructions, the robot uses its pneumatic injection probe to precisely deliver the designated fertilization unit to a predetermined three-dimensional location in the soil, thus completing a closed-loop, forward-looking, precise variable fertilization process.

[0078] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Any modifications, equivalent substitutions, or improvements made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method of fusing multi-modal data for precision variable rate fertilization of farmland soil, applied to farmland with slope micro-fissures, characterized in that, The method comprises the following steps: Collecting soil data and terrain data centered on slope micro-fissures and generating a spatial heterogeneity map reflecting the influence of the slope micro-fissures; Obtaining three-dimensional spatial distribution data of plant root systems and processing the three-dimensional spatial distribution data of the plant root systems by using a convolutional neural network to extract quantified root competition characteristics; Determining a nutrient absorption demand mode of multiple species according to the root competition characteristics and fusing the nutrient absorption demand mode with a three-dimensional voxel model representing a three-dimensional structure of soil to generate a three-dimensional water attribute distribution map; Based on the three-dimensional water attribute distribution map, combining climate data and a hydrology model to simulate water and fertilizer migration to predict a risk of excessive enrichment of nutrients in a downstream slope area; According to a path of the water and fertilizer migration simulation, using a soil physics model to optimize generation of differentiated fertilizer distribution parameters and mapping the differentiated fertilizer distribution parameters to the three-dimensional voxel grid to generate a precision variable fertilization scheme that clearly defines fertilizer types, application amounts and locations in three-dimensional space; and According to the three-dimensional voxel model adjustment scheme, controlling a variable fertilization device provided with a high-precision positioning system to accurately deliver a fertilization unit composed of at least one of fertilizer, water-retaining agent or soil conditioner to a three-dimensional spatial coordinate position of soil specified by the three-dimensional voxel model; The step of generating the spatial heterogeneity map comprises: Obtaining soil water and nutrient distribution data by deploying a soil sensor network around the slope micro-fissures; Combining digital elevation model to obtain terrain parameters including slope and aspect; Using a geographic information system to perform rasterization and superposition processing on the terrain parameters and the soil water and nutrient distribution data and statistically calculating a variation coefficient of nutrient concentration in the region to quantify heterogeneity; The step of determining a nutrient absorption demand mode of multiple species according to the root competition characteristics comprises: Fusing the root competition characteristics with soil type distribution data, adjusting a soil type data weight when the root competition characteristic value exceeds a preset threshold to calculate a multiple species competition intensity considering a rhizosphere environment and determining the nutrient absorption demand mode of the multiple species according to the multiple species competition intensity; The step of simulating water and fertilizer migration to predict a risk of excessive enrichment of nutrients comprises: Analyzing the three-dimensional water attribute distribution map to determine whether a root density of an upslope area is lower than a preset soil fixation capacity threshold; If the root density is lower than the soil fixation capacity threshold, fusing real-time monitoring data of climate with the hydrology model to generate a path of the water and fertilizer migration simulation; Comparing a total amount of predicted nutrient input of an end region of the path of the water and fertilizer migration simulation with a soil environmental capacity of the end region to quantify the risk of excessive enrichment of nutrients; The step of optimizing generation of differentiated fertilizer distribution parameters comprises: Using a soil physics model based on Richards equation to simulate movement of water and nutrients in a voxel after fertilization and combining a preset absorption rate of crops to fertilizer to maximize nutrient availability in a root zone and minimize a nutrient flux on a predicted loss path as an optimization objective to perform weighted calculation to optimize generation of the differentiated fertilizer distribution parameters.

2. The method of claim 1, wherein, The step of acquiring the three-dimensional spatial distribution data of the vegetation root system comprises: The root depth scanning data of the vegetation root system is acquired by using ground penetrating radar or root drilling method to form a root system distribution data set containing three-dimensional spatial position information.

3. The method of claim 1, wherein, The method further comprises: Based on the path of the water and fertilizer migration simulation, the farmland surface is automatically partitioned by using a clustering algorithm to divide the upper slope nutrient loss risk area and the lower slope nutrient excessive enrichment risk area; The automatic agricultural machinery linked with the partition result executes differentiated surface repair means for different partitions.

4. A precision variable rate fertilizer applicator for agricultural fields fusing multi-modal data, characterized by, It comprises: A data acquisition module is configured to execute the step of generating a spatial heterogeneity map in the method of claim 1; A model construction and analysis module is configured with a convolutional neural network, a hydrology model and a soil physics model, and is configured to execute the steps of extracting quantified root competition characteristics, generating a three-dimensional water property distribution map, simulating water and fertilizer migration, and generating a three-dimensional voxel model adjustment scheme in the method of claim 1; A precision fertilization execution module is configured to execute the step of accurately delivering a fertilization unit according to the three-dimensional voxel model adjustment scheme in the method of claim 1.

5. The apparatus of claim 4, wherein, The precision fertilization execution module is linked with the model construction and analysis module to convert the three-dimensional voxel model adjustment scheme into an execution instruction of a high-precision soil injection device or an automatic agricultural machinery.

Images