Edge end fruit tree fertilizer demand prediction method based on orchard intelligent water and fertilizer integration

Abstract

The invention relates to the technical field of intelligent agriculture, and particularly discloses an orchard intelligent water and fertilizer integration-based edge end fruit tree fertilizer demand prediction method, which comprises the following steps of: constructing each fruit tree as a graph node, and fusing canopy spatial characteristics extracted by unmanned aerial vehicle multispectral remote sensing and nutrient time sequence characteristics monitored by a soil sensor; space influence weights among nodes are calculated according to the terrain, the soil texture and the space distance, and a fruit tree individual relation graph is established; then, through a space-time diagram convolutional network, aggregating space neighborhood features and capturing time dynamic changes, and outputting target state features subjected to multi-scale enhancement; when the canopy conflicts with the soil characteristic indication, combined diagnosis is carried out based on a neighborhood consistency index and an adjacent node state, systematic nutrient stress and local non-nutrient factors are distinguished, and final fertilizer demand prediction is generated after the conflict is eliminated; a digital fertilization prescription map is generated through spatial interpolation, soil characteristics and slope correction, and the water and fertilizer integrated equipment is driven to execute precise operation.

Description

Technical Field

[0001] This invention relates to the field of smart agriculture technology, specifically to a method for predicting the fertilizer requirements of fruit trees at the edge of an intelligent water and fertilizer integration system in orchards. Background Technology

[0002] With the development of smart agriculture technology, orchard water and fertilizer integration management is gradually shifting from the traditional extensive model to precise intelligent control. In this process, accurately predicting the nutrient requirements of fruit trees is a core prerequisite for achieving precision fertilization. Existing technologies typically rely on canopy imagery obtained through drone remote sensing or soil sensor data deployed in the field to estimate nutrient requirements by establishing crop growth models or employing machine learning algorithms. In recent years, to further improve prediction accuracy, analytical methods integrating multi-source data have emerged. These methods attempt to combine spatial characteristics reflecting fruit tree canopy growth with temporal characteristics reflecting the root zone soil environment to obtain a more comprehensive basis for decision-making.

[0003] The existing technology has the following shortcomings:

[0004] The canopy spatial features (such as chlorophyll content) captured by UAV remote sensing are essentially a lagging reflection of the fruit trees' nutrient absorption and metabolism over a past period; while the nutrient data monitored by soil sensors reflects the current instantaneous nutrient supply status in the root zone. When fruit trees are in the transitional stage from nutrient sufficiency to nutrient stress, the system is highly susceptible to observing contradictory phenomena such as normal canopy indicators while soil nutrients have begun to decline, or canopy showing nutrient deficiency while soil data remains within the normal range. Most existing fusion technologies perform simple weighting or splicing at the feature layer or decision layer, lacking a mechanism to identify and arbitrate such contradictions caused by the inherent characteristics of plant physiological processes. This leads to model confusion and may even result in decisions that are less reliable than those using a single data source. For example, when local pests and diseases cause yellowing of leaves, fertilization may be mistakenly applied as a sign of nutrient deficiency, or the optimal fertilization opportunity may be missed in the early stages of systemic nutrient depletion due to still-available soil data. Summary of the Invention

[0005] The purpose of this invention is to provide a method for predicting the fertilizer requirements of fruit trees at the edge of an orchard based on intelligent water and fertilizer integration, so as to solve the problems mentioned above.

[0006] The objective of this invention can be achieved through the following technical solutions:

[0007] Based on the method for predicting the fertilizer requirements of fruit trees at the edge of the intelligent water and fertilizer integration in orchards, S1: Define each fruit tree in the orchard as a graph node; construct node features for each graph node, including the canopy spatial features of the fruit tree and the nutrient temporal features of the soil in the root zone of the fruit tree; calculate the spatial influence weight between any two node features, and construct the edges connecting each graph node according to the spatial influence weight to form a fruit tree individual relationship graph.

[0008] S2: Input the fruit tree individual relationship graph into the spatiotemporal graph convolutional network for processing; the spatiotemporal graph convolutional network aggregates the features of adjacent graph nodes along the edges in the spatial dimension, models the dynamic changes of the features of each graph node in the temporal dimension, and outputs the target state features of each graph node enhanced by spatiotemporal context information.

[0009] S3: For any target map node, when there is a conflict between the canopy spatial characteristics of the target map node and the fertilizer requirement status indicated by the root zone soil nutrient time series characteristics, a joint diagnosis is performed based on the target status characteristics of the adjacent map nodes of the target map node; based on the status of the adjacent map nodes, it is determined whether the conflict is caused by systemic nutrient stress or local non-nutrient factors, and the conflict is resolved to generate the final fertilizer requirement prediction of the target map node.

[0010] S4: Summarize the final fertilizer requirement predictions of all graph nodes, generate a fertilizer prescription map for the orchard, and distribute the fertilizer prescription map to the integrated water and fertilizer equipment to perform irrigation and fertilization operations.

[0011] As a further aspect of the present invention: the step of constructing node features for each graph node specifically includes:

[0012] We acquired UAV multispectral remote sensing images and root zone soil sensor time-series data under the same physiological time reference. We segmented the multispectral remote sensing images into individual fruit trees, extracted the spectral reflectance of the canopy for each segmented tree, and calculated the normalized vegetation index and chlorophyll spectral index, which together constituted the canopy spatial feature vector of the corresponding map node. We cleaned and interpolated the root zone soil sensor time-series data to form a soil nutrient content curve over a complete growth cycle, and extracted the mean, trend, and extreme point features of nutrient content from the corresponding curve to form the nutrient time-series feature vector of the corresponding map node. We then concatenated the canopy spatial feature vector and the nutrient time-series feature vector to generate node features that characterize the overall state of the fruit tree.

[0013] As a further aspect of the present invention: the formation of the fruit tree individual relationship diagram specifically includes:

[0014] Calculate the orchard's slope aspect and gradient to generate a surface water flow direction matrix. Based on this matrix, determine whether any two graph nodes have upstream and downstream hydraulic connections and calculate their hydraulic connection strength as the first weight component. Obtain the orchard's soil texture map; if two graph nodes have the same soil texture type, assign a fixed soil similarity bonus as the second weight component. Calculate the Euclidean distance between two graph nodes and generate a distance weight as the third weight component. Weight the first, second, and third weight components to obtain the final spatial influence weight. Establish undirected edges between all graph node pairs whose spatial influence weights exceed a preset threshold, thereby constructing a graph of individual fruit tree relationships.

[0015] As a further aspect of the present invention: the step of inputting the fruit tree individual relationship graph into a spatiotemporal graph convolutional network for processing specifically includes:

[0016] In each graph convolutional layer of the spatiotemporal graph convolutional network, the aggregation weights of the features of adjacent graph nodes are calculated. For each graph node, its features and the features of all adjacent graph nodes are used to calculate attention scores, which are determined based on the similarity between features and the spatial influence weights of the connecting edges. The features of each adjacent graph node are weighted and summed according to their corresponding attention scores to obtain the spatially aggregated features. In the temporal dimension, the feature sequences of graph nodes from multiple consecutive temporal sections are input into the gated recurrent unit. The retention and forgetting of historical information are controlled by the update gate and reset gate inside the unit to capture the dynamic evolution of the fruit tree's growth state. The spatially aggregated features and the temporal evolution features are fused to complete the deep extraction of spatiotemporal features.

[0017] As a further aspect of the present invention: the output of the target state features of each graph node enhanced with spatiotemporal context information specifically includes:

[0018] The graph node features output from the final layer of the spatiotemporal graph convolutional network are mapped to a latent space dimension that matches the fertilizer demand prediction task by passing them through a fully connected projection layer. Multi-scale features of graph nodes are extracted from the intermediate layers of the spatiotemporal graph convolutional network, and these multi-scale features are skip-connected with the features projected from the final layer to form an enhanced feature representation that includes local details and global semantics. The dimensions of the enhanced feature representation are recalibrated using the learned weight vectors to suppress noise dimensions and enhance key dimensions. The final generated target state features contain both the spatiotemporal evolution information of the graph nodes themselves and the topological influence of their neighborhood context.

[0019] As a further aspect of the present invention: the joint diagnosis based on the target state features of the target graph node's neighboring graph nodes specifically includes:

[0020] The similarity between the spatial features of the canopy and the temporal features of soil nutrients in the root zone of the target graph node and the corresponding features of each of its neighboring graph nodes is calculated to obtain the canopy feature similarity sequence and the soil feature similarity sequence. The two similarity sequences are weighted based on spatial influence weights to calculate the neighborhood consistency index of the target graph node in the canopy dimension and the soil dimension, respectively. A conflict quantification index is constructed. By comparing the difference between the neighborhood consistency index in the canopy dimension and the neighborhood consistency index in the soil dimension, and combining the nutrient stress intensity reflected in the target state features of neighboring graph nodes, a joint diagnostic feature vector for subsequent conflict type determination is generated.

[0021] As a further aspect of the present invention: the prediction of the final fertilizer requirement for generating target graph nodes specifically includes:

[0022] A threshold for determining systemic nutrient stress is set. When the neighborhood consistency index of the canopy dimension is higher than the threshold, it is defined as systemic nutrient stress, and the nutrient requirement state indicated by the spatial features of the canopy is adopted. A condition for determining local non-nutrient factors is set. When the neighborhood consistency index of the soil dimension is in the normal range and the canopy feature similarity sequence shows an outlier distribution, it is determined as a local non-nutrient factor, and the nutrient requirement state indicated by the temporal features of root zone soil nutrients is adopted. Based on the determination results, the corresponding nutrient requirement state is selected, and the final nutrient requirement prediction of the target graph node is generated by calibrating by introducing the spatial interpolation results of the nutrient requirement prediction of adjacent graph nodes.

[0023] As a further aspect of the present invention: the fertilizer prescription diagram for the fruit orchard specifically includes:

[0024] Based on the geographic coordinates of all graph nodes and the corresponding final fertilizer requirement predictions, a continuous fertilizer requirement distribution surface covering the entire orchard is generated. The fertilizer requirement distribution surface is overlaid with the soil type distribution map of the orchard for analysis. Differentiated fertilizer application correction coefficients are set according to different soil fertility retention characteristics to adjust the distribution surface. Multiple fertilization management zones are divided according to slope level, and the fertilizer application rate is normalized in each zone to ensure the uniformity of fertilization operations within the same zone. The processed fertilizer requirement distribution data is converted into a standard prescription map file format to generate a digital fertilization prescription map containing geographic coordinate references and zone fertilization instructions.

[0025] The beneficial effects of this invention are:

[0026] (1) This invention constructs an individual relationship graph of fruit trees, utilizes a spatiotemporal graph convolutional network to deeply integrate canopy spatial features and soil nutrient temporal features, and fully considers the spatial correlation and temporal dynamics among fruit trees. When data conflicts occur, a joint diagnostic mechanism based on the neighborhood consistency index is innovatively introduced, which can intelligently determine whether the root cause of the anomaly is a wide-area nutrient stress or a local non-nutrient factor, thereby making a decision closer to the actual needs. This diagnostic method based on swarm intelligence and spatiotemporal context reduces the sensitivity to isolated abnormal data, making the fertilizer requirement prediction results more robust and reliable.

[0027] (2) The fertilizer prescription map generated by this invention is not simply a spatial interpolation of the predicted values, but rather a multi-round optimization that integrates the soil characteristics and topographic information of the orchard. By using the soil fertility retention characteristic correction coefficient, the amount of fertilizer is adjusted for different textures such as sandy soil, loam, and clay, so that the fertilizer supply matches the soil's retention capacity. Through slope zoning management, the amount of fertilizer is reduced in steep slope areas to reduce the risk of runoff pollution, while the uniformity of fertilization is optimized in gentle slope areas. The final prescription map can guide the integrated water and fertilizer equipment to perform precise operations of "distribution on demand and fertilization according to local conditions", avoiding fertilizer waste and local deficiencies caused by traditional uniform fertilization, thereby achieving multiple goals of saving water and fertilizer, improving quality and efficiency, and protecting the environment. Attached Figure Description

[0028] The invention will now be further described with reference to the accompanying drawings.

[0029] Figure 1 This is a flowchart of the method of the present invention. Detailed Implementation

[0030] 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.

[0031] Please see Figure 1 As shown, this invention is a method for predicting the nutrient requirements of fruit trees at the edge of an orchard based on intelligent fertigation, comprising the following steps:

[0032] S1: Define each fruit tree in the orchard as a graph node; construct node features for each graph node, including the canopy spatial features of the fruit tree and the nutrient temporal features of the soil in the root zone of the fruit tree; calculate the spatial influence weight between any two node features, and construct the edges connecting each graph node according to the spatial influence weight to form a fruit tree individual relationship graph.

[0033] S2: Input the fruit tree individual relationship graph into the spatiotemporal graph convolutional network for processing; the spatiotemporal graph convolutional network aggregates the features of adjacent graph nodes along the edges in the spatial dimension, models the dynamic changes of the features of each graph node in the temporal dimension, and outputs the target state features of each graph node enhanced by spatiotemporal context information.

[0034] S3: For any target map node, when there is a conflict between the canopy spatial characteristics of the target map node and the fertilizer requirement status indicated by the root zone soil nutrient time series characteristics, a joint diagnosis is performed based on the target status characteristics of the adjacent map nodes of the target map node; based on the status of the adjacent map nodes, it is determined whether the conflict is caused by systemic nutrient stress or local non-nutrient factors, and the conflict is resolved to generate the final fertilizer requirement prediction of the target map node.

[0035] S4: Summarize the final fertilizer requirement predictions of all graph nodes, generate a fertilizer prescription map for the orchard, and distribute the fertilizer prescription map to the integrated water and fertilizer equipment to perform irrigation and fertilization operations.

[0036] In S1, each fruit tree in the orchard is defined as a graph node; node features are constructed for each graph node, including the canopy spatial features and the nutrient temporal features of the soil in the root zone; the spatial influence weight between any two node features is calculated, and edges connecting each graph node are constructed based on the spatial influence weight to form an individual fruit tree relationship graph, specifically including:

[0037] Each independently growing fruit tree in the orchard is defined as a graph node, serving as the basic unit for the subsequent graph structure model. To construct node features, multi-source data with synchronized time and a unified reference standard is required. First, a multi-rotor UAV equipped with a multispectral camera is operated under clear, cloudless weather conditions with a solar altitude angle greater than 45 degrees, flying at a predetermined altitude above the fruit tree canopy to acquire high-resolution multispectral remote sensing images covering the entire orchard. The image bands include at least red, green, blue, near-infrared, and red-edge bands. Simultaneously, a soil sensor network deployed within the root zone of each fruit tree synchronously records minute-level time-series data on the nitrogen, phosphorus, and potassium content in the soil, as well as the soil volumetric water content. To ensure the consistency of the data time reference, the precise time point of the UAV aerial photography is recorded, and this time point is used to align with the time-series data recorded by the soil sensors. All data are marked with a unified physiological timestamp, which is determined by the phenological stage and effective accumulated temperature of the fruit tree.

[0038] The construction of node feature vectors includes the extraction of canopy spatial feature vectors and the extraction of root zone soil nutrient temporal feature vectors. For the extraction of canopy spatial feature vectors, a deep learning-based instance segmentation algorithm is first used to process the acquired multispectral remote sensing imagery, identifying and segmenting the individual canopy outline of each fruit tree. For each segmented individual canopy, the average spectral reflectance of all pixels in each band is calculated. Based on these average reflectance values, the normalized vegetation index (NVI) is calculated, which is obtained by dividing the difference in reflectance between the near-infrared and red bands by the sum of the two. Simultaneously, the chlorophyll spectral index is calculated, which is typically derived by combining the reflectance of the red-edge band and the near-infrared band, which are sensitive to chlorophyll content. Finally, the calculated normalized vegetation index, chlorophyll spectral index, and other relevant spectral indices together constitute the canopy spatial feature vector of the graph node.

[0039] To extract the temporal feature vector of soil nutrients in the root zone, the raw time-series data collected by the soil sensor network is first cleaned to remove obvious outliers caused by sensor malfunctions or signal interference. Then, a time-series-based linear interpolation method is used to fill in transient missing values, thus forming a complete and continuous soil nutrient content curve for the fruit tree's growth cycle. From this processed nutrient content curve, statistical features characterizing the soil nutrient status are extracted. These features include: the arithmetic mean of nutrient content within the growth cycle, representing the overall level; the slope obtained by linear fitting the curve, representing the continuous consumption or accumulation trend of nutrients; and the maximum and minimum values ​​of nutrient content within the cycle, representing extreme fluctuations in nutrient levels. The above-mentioned mean, trend, and extreme point features together constitute the nutrient temporal feature vector of the node in the graph. Finally, the canopy spatial feature vector and the nutrient temporal feature vector are concatenated along the feature dimension to generate a comprehensive node feature vector, used to characterize the overall growth and nutritional status of the fruit tree.

[0040] The calculation of spatial influence weights relies on the orchard's topographic and environmental attribute data. First, based on high-precision digital elevation model data of the orchard, the slope aspect and gradient of each grid cell are calculated, and a specific algorithm is used to generate a surface water flow direction matrix. This matrix characterizes the flow direction of surface runoff after rainfall or irrigation. Based on this surface water flow direction matrix, the positional relationship between any two graph nodes representing fruit trees is determined: if a continuous hydraulic path exists from one fruit tree to another, an upstream-downstream hydraulic connection is established. The hydraulic connection strength is calculated based on the cumulative catchment area of ​​the upstream node to the downstream node and the slope between them; this strength serves as the first weight component of the spatial influence weights.

[0041] A soil texture map of the orchard is obtained, which is drawn through soil sampling and laboratory analysis. If the soil texture of the fruit tree growing area corresponding to any two map nodes is classified into the same type, a fixed soil similarity score is assigned to that node pair. This score serves as the second weight component of the spatial influence weight. Its theoretical basis is that soil textures usually have similar fertilizer retention characteristics and nutrient transport patterns.

[0042] Calculate the Euclidean distance between the fruit trees represented by two graph nodes in terms of geographical location. Based on the principle of distance decay, apply a monotonically decreasing function to map the Euclidean distance into a distance weight. This weight serves as the third weight component of the spatial influence weight to simulate the spatial proximity effect.

[0043] After determining the first, second, and third weight components, a weighted summation method is used to superimpose these three components, calculating the final spatial influence weight between any two graph nodes. An undirected edge is established between all graph node pairs whose calculated spatial influence weights exceed a pre-set empirical threshold, thus connecting the discrete graph nodes into a network topology and ultimately completing the construction of the fruit tree individual relationship graph. This graph structure quantifies the mutual influence relationships between fruit tree individuals in terms of physical space and environmental attributes.

[0044] In S2, the fruit tree individual relationship graph is input into a spatiotemporal graph convolutional network for processing. The spatiotemporal graph convolutional network aggregates features of adjacent graph nodes along the edges in the spatial dimension, models the dynamic changes of each graph node's features in the temporal dimension, and outputs the target state features of each graph node enhanced with spatiotemporal context information, specifically including:

[0045] The fruit tree individual relationship graph constructed in the aforementioned steps is input into a spatiotemporal graph convolutional network for processing. The network's input data consists of two components: first, a node feature matrix composed of the feature vectors of all graph nodes, where each row corresponds to the comprehensive feature vector of a graph node; second, an adjacency matrix describing the connections between graph nodes, with elements filled with previously calculated spatial influence weights. Before inputting into the network, the node feature matrix is ​​standardized by calculating the mean and standard deviation of all nodes for each feature dimension, then subtracting the mean from each feature value and dividing by the standard deviation, making the data distribution for each feature dimension approximate a standard normal distribution. This preprocessing step helps improve the stability and convergence speed of subsequent model training. The adjacency matrix is ​​also normalized, typically using a symmetric normalization method, to make the weight distribution during feature aggregation more reasonable.

[0046] In each graph convolutional layer of the spatiotemporal graph convolutional network, the core operation is to compute and aggregate the feature information of neighboring graph nodes. For the currently processed central graph node, it is first necessary to calculate the feature aggregation weights between it and all directly connected neighboring graph nodes. This computation process is implemented through an attention mechanism, specifically, calculating the similarity between the feature vectors of the central graph node and each neighboring graph node. This similarity is obtained by calculating the inner product of the two feature vectors and then applying the softmax function for normalization to obtain an initial attention score. At the same time, the spatial influence weight of the edge connecting the central node and the neighboring nodes is also taken into consideration. The final comprehensive attention score is determined by a weighted combination of the feature similarity score and the spatial influence weight, where the introduction of the spatial influence weight gives more importance to neighboring nodes that have a greater influence in physical space or environmental attributes during the feature aggregation process. After determining the comprehensive attention score of all neighboring nodes relative to the central node, the feature vector of each neighboring node is multiplied by its corresponding attention score, and then all these weighted feature vectors are summed to obtain a new feature representation of the central node after spatial neighborhood information enhancement, i.e., the spatially aggregated feature.

[0047] In the time dimension, it is necessary to capture the dynamic changes of graph node features over time. The graph node feature sequences from multiple consecutive acquisition time points are organized chronologically to form a time-step sequence data. This sequence data is input into a gated recurrent unit (GRU) for temporal modeling. The GRU contains two gating mechanisms: an update gate and a reset gate. The update gate controls how much state information from the previous time step is retained in the current time step. It is calculated by applying a sigmoid activation function after a linear transformation of the current input features and the hidden state from the previous time step, outputting a value between 0 and 1. The reset gate determines how much past information needs to be forgotten; its calculation method is similar to the update gate but with different parameters. The candidate hidden state at the current time step is calculated jointly by the forgotten hidden state from the previous time step controlled by the reset gate and the current input features. Finally, the hidden state at the current time step is composed of the portion of the hidden state from the previous time step retained (controlled by the update gate) and the portion of the candidate hidden state accepted. Through this gating mechanism, the gated recurrent unit can adaptively learn the dynamic evolution patterns of characteristics during fruit tree growth, capturing temporal patterns such as seasonal changes in nutrient requirements and the gradual process of canopy growth.

[0048] After obtaining the spatial aggregated features and temporal evolutionary features respectively, it is necessary to effectively fuse these two different sources of features. Common fusion strategies include feature concatenation, weighted summation, or nonlinear combination through additional neural network layers. Feature concatenation directly connects the spatial aggregated feature vector and the temporal evolutionary feature vector along the feature dimension to form a more expressive fused feature vector. Weighted summation assigns learnable weights to the two features and then performs element-wise weighted addition. Regardless of the fusion strategy used, the fused features will be passed through a nonlinear activation function, such as the ReLU function, to introduce nonlinear transformation capabilities. This spatiotemporal feature fusion process may be repeated in multiple layers of the spatiotemporal graph convolutional network. Each layer further extracts higher-level spatiotemporal feature patterns based on the more abstract feature representation provided by the previous layer, thereby achieving deep extraction of spatiotemporal features.

[0049] The graph node features output from the final layer of the spatiotemporal graph convolutional network (SPCB) first pass through a fully connected projection layer, which maps the feature dimensions to a latent space dimension that matches the subsequent demand prediction task. Simultaneously, multi-scale features of the graph nodes are extracted from the intermediate layers of the SPCB. These intermediate layer features typically retain more local detail information, while the final layer features contain richer global semantic information. Skip connections are used to combine these features from different layers. Specifically, the intermediate layer features are appropriately dimensionalized and then summed or concatenated with the final layer projection features, forming an enhanced feature representation containing multi-scale information.

[0050] To further optimize the quality of the enhanced feature representation, an adaptive feature selection mechanism is introduced. This mechanism learns a weight vector with the same dimension as the enhanced feature representation, where each element corresponds to a weight for a feature dimension. These weights are automatically generated by analyzing the overall distribution characteristics of the features. Typically, global average pooling is used to compress the information of each feature dimension across all nodes into a single statistic. The final weight vector is then generated through a fully connected layer with a sigmoid activation function. This weight vector is then multiplied element-wise with the enhanced feature representation to recalibrate each feature dimension, thereby suppressing noisy dimensions with limited information while enhancing the key feature dimensions for fertilizer demand prediction. The resulting target state features incorporate both the spatiotemporal evolution information of the graph nodes themselves and the topological influence of their neighborhood context, providing high-quality feature input for subsequent fertilizer demand decisions.

[0051] In S3, for any target map node, when there is a conflict between the canopy spatial characteristics of the target map node and the nutrient demand status indicated by the root zone soil nutrient time series characteristics, a joint diagnosis is performed based on the target state characteristics of the adjacent map nodes. Based on the status of the adjacent map nodes, it is determined whether the conflict is caused by systemic nutrient stress or local non-nutrient factors, and the conflict is resolved to generate the final nutrient demand prediction for the target map node. Specifically, this includes:

[0052] When the system detects an inconsistency between the nutrient requirement status indicated by the canopy spatial features of a target graph node and the nutrient requirement status indicated by the temporal features of soil nutrients in its root zone, a joint diagnostic process is initiated. First, cross-modal feature similarity calculation is performed. The canopy spatial feature vector of the target graph node is compared with the canopy spatial feature vectors of each of its neighboring graph nodes using a cosine similarity method. This method calculates the cosine of the angle between the two vectors in multidimensional space, with the result ranging from -1 to 1. The closer the value is to 1, the more consistent the directions of the two vectors and the more similar their features. Similarly, the temporal feature vector of soil nutrients in the root zone of the target graph node is also compared with the corresponding feature vector of each of its neighboring graph nodes using the same cosine similarity method. Through these calculations, two ordered similarity sequences are obtained: one is the sequence of similarity between the target node and its neighboring nodes in terms of canopy spatial features, called the canopy feature similarity sequence; the other is the sequence of similarity between the target node and its neighboring nodes in terms of soil nutrient temporal features, called the soil feature similarity sequence. These two sequences reflect the degree of consistency between the target node and its neighboring nodes in terms of canopy performance and soil conditions from different modalities.

[0053] After obtaining the initial similarity sequence, the influence of spatial heterogeneity needs to be considered. Therefore, the spatial influence weights calculated when constructing the fruit tree individual relationship map are introduced to correct the similarity sequence. For the canopy feature similarity sequence, each similarity value in the sequence is multiplied by its corresponding spatial influence weight to obtain a weighted canopy feature similarity value; similarly, the soil feature similarity sequence undergoes the same weighting process. Then, the arithmetic mean of the two weighted similarity sequences is calculated to obtain two comprehensive indices: the canopy-dimensional neighborhood consistency index and the soil-dimensional neighborhood consistency index. The calculation of both indices considers the spatial relationship strength between neighboring nodes and the target node; neighboring nodes with larger spatial influence weights contribute more to the final index value. The canopy-dimensional neighborhood consistency index reflects the consistency of fruit tree canopy performance within the local area where the target node is located, while the soil-dimensional neighborhood consistency index reflects the consistency of soil nutrient conditions within that area. These two indices provide a quantitative basis for subsequent conflict type judgment.

[0054] To systematically analyze and quantify the characteristic conflicts existing in target nodes, it is necessary to construct specialized conflict quantification indicators. First, the absolute difference between the canopy-dimensional neighborhood consistency index and the soil-dimensional neighborhood consistency index is calculated. This difference reflects the degree of deviation between the two modal characteristics in terms of spatial consistency. Simultaneously, the nutrient stress intensity reflected in the target state characteristics of adjacent graph nodes is analyzed. This is typically determined by analyzing the numerical distribution of nutrient stress-related dimensions in the feature vectors of adjacent nodes. Based on the above analysis results, a joint diagnostic feature vector is constructed. This vector contains diagnostic information in multiple dimensions: the original values ​​of the canopy-dimensional neighborhood consistency index and the soil-dimensional neighborhood consistency index, the difference between the two consistency indices, the statistical characteristics of the nutrient stress intensity of adjacent nodes (such as mean, maximum, and coefficient of variation), and the original degree of difference between the two modal characteristics of the target node itself. This joint diagnostic feature vector integrates the intensity information of the target node's own characteristic conflict, neighborhood consistency pattern, and neighborhood nutrient status, providing comprehensive input information for subsequent conflict type determination.

[0055] A threshold for determining systemic nutrient stress is set, determined through historical data analysis and expert experience, typically between 0.7 and 0.9, indicating a high degree of neighborhood consistency is required to classify it as systemic stress. When the neighborhood consistency index in the canopy dimension exceeds the set threshold for systemic nutrient stress, and the nutrient stress intensity reflected in the target state features of adjacent graph nodes is generally high, the current conflict is determined to be caused by systemic nutrient stress. This indicates that the entire area where the target node is located faces a similar nutrient deficiency problem, and the canopy nutrient deficiency symptoms exhibited by the target node reflect the true nutrient demand. Under this determination, the system adopts the nutrient requirement status indicated by the canopy spatial features as the decision basis, assuming that soil nutrient monitoring results may not accurately reflect the actual nutrient needs of plants due to measurement timing or location, thus prioritizing the satisfaction of crop physiological needs.

[0056] The system establishes criteria for identifying local non-nutrient factors, including that the soil dimensional neighborhood consistency index must be within the normal range. This normal range is determined through statistical analysis of data from a large number of normally growing orchards and is typically between 0.6 and 0.8. Simultaneously, the canopy feature similarity sequence must exhibit a significant outlier distribution, meaning the target node differs significantly from most of its neighboring nodes in canopy features. Outlier distribution is determined using statistical methods, such as calculating the interquartile range (IQM) of the canopy feature similarity sequence. Values ​​below the first IQM minus 1.5 IQM or above the third IQM plus 1.5 IQM are considered outliers. When all the above conditions are met, the current conflict is determined to be caused by local non-nutrient factors, such as pest or disease infection, mechanical damage, or localized moisture abnormalities. Under this determination, the system adopts the nutrient requirement status indicated by the root zone soil nutrient time series characteristics as the decision-making basis, assuming that canopy anomalies are caused by non-nutrient factors, and that soil nutrient monitoring results better reflect the true nutrient supply situation.

[0057] After selecting the appropriate fertilizer requirement state based on the aforementioned judgment results, spatial interpolation results of fertilizer requirement predictions for adjacent graph nodes are introduced for calibration to further improve prediction accuracy. An inverse distance weighted spatial interpolation method is used, calculating weights based on the spatial distance between the target node and each adjacent node; the closer the adjacent node, the greater its weight in the predicted fertilizer requirement value. The selected fertilizer requirement state and the spatial interpolation results are then weighted and combined to generate the final fertilizer requirement prediction for the target graph node. This calibration process effectively utilizes the principle of spatial autocorrelation, ensuring that the prediction results consider both the characteristics of the target node itself and the spatial trends of its neighborhood, thereby improving the spatial continuity and stability of the prediction while maintaining the ability to respond to local anomalies. Finally, the system outputs the final fertilizer requirement prediction for each graph node, providing a reliable data foundation for generating precise fertilization prescriptions.

[0058] In S4, the final fertilizer requirement predictions for all graph nodes are summarized, a fertilizer prescription map for the orchard is generated, and the fertilizer prescription map is sent to the integrated water and fertilizer system to perform irrigation and fertilization operations. Specifically, this includes:

[0059] Based on the geographic coordinates of all graph nodes and their corresponding predicted final nutrient requirements, a spatial interpolation algorithm is used to generate a continuous nutrient requirement distribution surface covering the entire orchard. The geographic coordinates of each graph node are precisely obtained through global satellite navigation system positioning, with a planar coordinate accuracy better than 0.05 meters. The spatial interpolation process employs the ordinary kriging method, which fully considers the spatial autocorrelation of data points and provides an estimate of the interpolation error. Specifically, the spatial variation structure of the predicted nutrient requirements of all graph nodes is first analyzed. The semivariogram values ​​at different distances are calculated using an experimental variogram function, and a theoretical variogram model is fitted. Based on this model, the spatial weights between each grid cell location to be interpolated and its surrounding known graph nodes are calculated. These weights minimize the estimated variance of the interpolation result. Finally, the estimated nutrient requirement of each grid cell is obtained by weighted summation of the predicted nutrient requirements of the surrounding known graph nodes, thus forming a spatially continuous nutrient requirement distribution surface. The spatial resolution of this surface is set to 1 meter × 1 meter according to the operational accuracy of the integrated water and fertilizer system, ensuring that the generated prescription map meets the actual operational requirements of precision fertilization.

[0060] After generating the initial nutrient requirement distribution surface, the fertilization rate needs to be finely adjusted by combining it with the orchard's soil type distribution map. The soil type distribution map, drawn from a previous soil survey, details the soil texture categories in different areas of the orchard, such as sandy soil, loam, and clay. Different soil textures exhibit significant differences in their ability to retain fertilizer. Sandy soils, due to their poor fertilizer retention capacity, require an increase in fertilization frequency or amount; while clay soils, with their strong fertilizer retention capacity, allow for a reduction in the amount of fertilizer per unit area. Therefore, a specific fertilization rate correction coefficient is assigned to each soil texture type. These coefficients are determined based on long-term field trial data: the correction coefficient for sandy soil ranges from 1.1 to 1.3, for clay soil from 0.7 to 0.9, and for loam is set to 1.0. The nutrient requirement distribution surface is then overlaid with the soil type distribution map. The estimated nutrient requirement for each soil type area is multiplied by the corresponding correction coefficient to generate a nutrient requirement distribution surface corrected for soil characteristics. This revision process ensures that fertilization recommendations can be adapted to the nutrient retention characteristics of different soil conditions, thereby improving fertilizer utilization efficiency.

[0061] Considering the impact of terrain slope on the uniformity of fertilization operations and the risk of fertilizer runoff, fertilization management zones need to be divided according to slope grade. Based on orchard digital elevation model data, the slope value of each grid cell is calculated, dividing the orchard into three slope grades: 0 to 5 degrees for gentle slopes, 5 to 15 degrees for medium slopes, and above 15 degrees for steep slopes. Within each slope grade zone, the fertilizer requirement, adjusted for soil characteristics, is normalized. The purpose of normalization is to ensure that the fertilizer application rate remains relatively consistent within the same fertilization management zone, facilitating mechanized fertilization operations. The process first calculates the statistical distribution characteristics of fertilizer requirement within each slope zone, including the mean and standard deviation. Then, the Z-score standardization method is used to adjust the fertilizer requirement values ​​within the zone to a relatively uniform level, while preserving the differences in fertilizer requirement between different zones. For steep slope areas, an additional runoff risk reduction coefficient is applied to appropriately reduce the fertilizer application rate per unit area, thereby reducing the risk of fertilizer loss and environmental pollution caused by surface runoff.

[0062] The obtained final fertilizer requirement distribution data is converted into a standardized prescription map file that meets the requirements of agricultural machinery operations. This file is organized in XML format and includes a header and a data body. The header records metadata about the prescription map, including the orchard name, geographic coordinate system parameters, generation time, data version number, and type of operating equipment. The data body organizes fertilizer application instructions in a grid structure. Each grid cell contains the latitude and longitude coordinates of its center point, the target fertilizer application amount, and a status flag for the fertilization operation. The target fertilizer application amount is expressed in kilograms per hectare, with precision to one decimal place. The status flag indicates whether the grid cell is in an operable state, preventing fertilization operations in non-operational areas such as buildings and water bodies within the orchard. The generated digital fertilizer prescription map can be transmitted to the control system of the integrated water and fertilizer system via wireless network or mobile storage device to guide the precise execution of variable fertilizer application operations.

[0063] The working principle of this invention is as follows: Each fruit tree in the orchard is defined as a graph node. Canopy spatial features and root zone soil nutrient temporal features are simultaneously collected using UAV multispectral remote sensing and soil sensors to construct a node feature vector that integrates spectral indices and nutrient statistical features. Based on a digital elevation model, soil quality map, and spatial distance, the spatial influence weights between nodes are calculated to construct an individual fruit tree relationship graph. Next, the relationship graph is input into a spatiotemporal graph convolutional network. Spatial neighborhood features are aggregated through an attention mechanism, and time-varying dynamics are modeled using gated recurrent units. After fusing spatiotemporal features, multi-scale feature enhancement and adaptive recalibration are performed to generate target state features containing spatiotemporal context information. Then, when a node experiences a conflict between canopy and soil features, a neighborhood consistency index is calculated. The root cause of the conflict is determined by combining the states of adjacent nodes, distinguishing between systemic nutrient stress and local non-nutrient factors. Spatial interpolation calibration is used to generate the final nutrient requirement prediction. Finally, based on all node predictions, Kriging interpolation is used to generate a nutrient requirement distribution surface. This surface is then corrected and optimized by combining soil characteristics and slope zoning, and converted into a standard prescription map format before being sent to an integrated water and fertilizer system for precision fertilization.

[0064] The foregoing has provided a detailed description of one embodiment of the present invention, but this description is merely a preferred embodiment and should not be construed as limiting the scope of the invention. All equivalent variations and modifications made within the scope of the claims of this invention should still fall within the patent coverage of this invention.

Claims

1. A method for predicting the nutrient requirements of fruit trees at the edge of an intelligent fertigation system in orchards, characterized in that: Includes the following steps: S1: Define each fruit tree in the orchard as a graph node; construct node features for each graph node, including the canopy space features of the fruit tree and the nutrient temporal features of the soil in the root zone of the fruit tree; Calculate the spatial influence weight between any two node features, and construct the edges connecting each node in the graph based on the spatial influence weight to form a fruit tree individual relationship graph; S2: Input the fruit tree individual relationship graph into the spatiotemporal graph convolutional network for processing; the spatiotemporal graph convolutional network aggregates the features of adjacent graph nodes along the edges in the spatial dimension, models the dynamic changes of the features of each graph node in the temporal dimension, and outputs the target state features of each graph node enhanced by spatiotemporal context information. S3: For any target map node, when there is a conflict between the canopy spatial characteristics of the target map node and the fertilizer requirement status indicated by the root zone soil nutrient time series characteristics, a joint diagnosis is performed based on the target status characteristics of the adjacent map nodes of the target map node; based on the status of the adjacent map nodes, it is determined whether the conflict is caused by systemic nutrient stress or local non-nutrient factors, and the conflict is resolved to generate the final fertilizer requirement prediction of the target map node. S4: Summarize the final fertilizer requirement predictions of all graph nodes, generate a fertilizer prescription map for the orchard, and distribute the fertilizer prescription map to the integrated water and fertilizer equipment to perform irrigation and fertilization operations.

2. The method for predicting the nutrient requirements of fruit trees at the edge of an orchard based on intelligent fertigation as described in claim 1, characterized in that, The construction of node features for each graph node specifically includes: We acquired UAV multispectral remote sensing images and root zone soil sensor time-series data under the same physiological time reference. We segmented the multispectral remote sensing images into individual fruit trees, extracted the spectral reflectance of the canopy for each segmented tree, and calculated the normalized vegetation index and chlorophyll spectral index, which together constituted the canopy spatial feature vector of the corresponding map node. We cleaned and interpolated the root zone soil sensor time-series data to form a soil nutrient content curve over a complete growth cycle, and extracted the mean, trend, and extreme point features of nutrient content from the corresponding curve to form the nutrient time-series feature vector of the corresponding map node. We then concatenated the canopy spatial feature vector and the nutrient time-series feature vector to generate node features that characterize the overall state of the fruit tree.

3. The method for predicting the nutrient requirements of fruit trees at the edge of an orchard based on intelligent water and fertilizer integration as described in claim 1, characterized in that, The formation of the fruit tree individual relationship diagram specifically includes: Calculate the orchard's slope aspect and gradient to generate a surface water flow direction matrix. Based on this matrix, determine whether any two graph nodes have upstream and downstream hydraulic connections and calculate their hydraulic connection strength as the first weight component. Obtain the orchard's soil texture map; if two graph nodes have the same soil texture type, assign a fixed soil similarity bonus as the second weight component. Calculate the Euclidean distance between two graph nodes and generate a distance weight as the third weight component. Weight the first, second, and third weight components to obtain the final spatial influence weight. Establish undirected edges between all graph node pairs whose spatial influence weights exceed a preset threshold, thereby constructing a graph of individual fruit tree relationships.

4. The method for predicting the nutrient requirements of fruit trees at the edge of an orchard based on intelligent water and fertilizer integration as described in claim 1, characterized in that, The process of inputting the fruit tree individual relationship graph into a spatiotemporal graph convolutional network for processing specifically includes: In each graph convolutional layer of the spatiotemporal graph convolutional network, the aggregation weights of the features of adjacent graph nodes are calculated. For each graph node, its features and the features of all adjacent graph nodes are used to calculate attention scores, which are determined based on the similarity between features and the spatial influence weights of the connecting edges. The features of each adjacent graph node are weighted and summed according to their corresponding attention scores to obtain the spatially aggregated features. In the temporal dimension, the feature sequences of graph nodes from multiple consecutive temporal sections are input into the gated recurrent unit. The retention and forgetting of historical information are controlled by the update gate and reset gate inside the unit to capture the dynamic evolution of the fruit tree's growth state. The spatially aggregated features and the temporal evolution features are fused to complete the deep extraction of spatiotemporal features.

5. The method for predicting the nutrient requirements of fruit trees at the edge of an orchard based on intelligent fertigation as described in claim 1, characterized in that, The output of the target state features of each graph node enhanced with spatiotemporal context information specifically includes: The graph node features output from the final layer of the spatiotemporal graph convolutional network are mapped to a latent space dimension that matches the fertilizer demand prediction task by passing them through a fully connected projection layer. Multi-scale features of graph nodes are extracted from the intermediate layers of the spatiotemporal graph convolutional network, and these multi-scale features are skip-connected with the features projected from the final layer to form an enhanced feature representation that includes local details and global semantics. The dimensions of the enhanced feature representation are recalibrated using the learned weight vectors to suppress noise dimensions and enhance key dimensions. The final generated target state features contain both the spatiotemporal evolution information of the graph nodes themselves and the topological influence of their neighborhood context.

6. The method for predicting the nutrient requirements of fruit trees at the edge of an orchard based on intelligent fertigation as described in claim 1, characterized in that, The joint diagnosis based on the target state features of adjacent graph nodes of the target graph node specifically includes: The similarity between the spatial features of the canopy and the temporal features of soil nutrients in the root zone of the target graph node and the corresponding features of each of its neighboring graph nodes is calculated to obtain the canopy feature similarity sequence and the soil feature similarity sequence. The two similarity sequences are weighted based on spatial influence weights to calculate the neighborhood consistency index of the target graph node in the canopy dimension and the soil dimension, respectively. A conflict quantification index is constructed. By comparing the difference between the neighborhood consistency index in the canopy dimension and the neighborhood consistency index in the soil dimension, and combining the nutrient stress intensity reflected in the target state features of neighboring graph nodes, a joint diagnostic feature vector for subsequent conflict type determination is generated.

7. The method for predicting the nutrient requirements of fruit trees at the edge of an orchard based on intelligent water and fertilizer integration as described in claim 1, characterized in that, The final fertilizer requirement prediction for the generated target graph nodes specifically includes: A threshold for determining systemic nutrient stress is set. When the neighborhood consistency index of the canopy dimension is higher than the threshold, it is defined as systemic nutrient stress, and the nutrient requirement state indicated by the spatial features of the canopy is adopted. A condition for determining local non-nutrient factors is set. When the neighborhood consistency index of the soil dimension is in the normal range and the canopy feature similarity sequence shows an outlier distribution, it is determined as a local non-nutrient factor, and the nutrient requirement state indicated by the temporal features of root zone soil nutrients is adopted. Based on the determination results, the corresponding nutrient requirement state is selected, and the final nutrient requirement prediction of the target graph node is generated by calibrating by introducing the spatial interpolation results of the nutrient requirement prediction of adjacent graph nodes.

8. The method for predicting the nutrient requirements of fruit trees at the edge of an orchard based on intelligent water and fertilizer integration as described in claim 1, characterized in that, The fertilizer prescription diagram for the fruit orchard specifically includes: Based on the geographic coordinates of all graph nodes and the corresponding final fertilizer requirement predictions, a continuous fertilizer requirement distribution surface covering the entire orchard is generated. The fertilizer requirement distribution surface is overlaid with the soil type distribution map of the orchard for analysis. Differentiated fertilizer application correction coefficients are set according to different soil fertility retention characteristics to adjust the distribution surface. Multiple fertilization management zones are divided according to slope level, and the fertilizer application rate is normalized in each zone to ensure the uniformity of fertilization operations within the same zone. The processed fertilizer requirement distribution data is converted into a standard prescription map file format to generate a digital fertilization prescription map containing geographic coordinate references and zone fertilization instructions.

Images