Juicy peach yield prediction method based on comprehensive data analysis

Abstract

The invention discloses a juicy peach yield prediction method based on comprehensive data analysis, and particularly relates to the technical field of agricultural intelligent perception. The method comprises the following steps: collecting multi-source data of an orchard, and constructing a data set containing meteorological, physiological, remote sensing and soil information; fruit tree physiological response parameters are extracted, and a bimodal diagram structure fusing the spatial adjacency relation and the physiological state similarity is established in combination with the dynamic climate anomaly index; inputting the graph structure into a graph neural network model, extracting spatial-temporal characteristics, dynamically adjusting an edge weight and a propagation layer number, and constructing an adaptive model; performing region division and weighted summarization according to a model output result, and finally obtaining a predicted value of the total yield of the orchard; the method improves the prediction accuracy under the conditions of complex climate and unstable data, and is suitable for refined orchard management.

Description

Technical Field

[0001] This invention relates to the field of agricultural intelligent sensing technology, specifically to a method for predicting peach yield based on comprehensive data analysis. Background Technology

[0002] In recent years, with the rapid development of smart agriculture, precision management and yield prediction in fruit tree cultivation have become research hotspots. As a fruit tree with high economic benefits and sensitivity to climate and soil conditions, the yield of peaches is influenced by a combination of factors, including meteorological factors (temperature, precipitation, sunshine, etc.), soil nutrient status, pest and disease severity, varietal differences, pruning techniques, and pollination rate during flowering.

[0003] However, current technologies for predicting peach yield mainly rely on single-variable models or manual judgment based on empirical rules, which are insufficient to handle nonlinear changes caused by the interaction of multiple factors. For example, in peach orchards in mountainous and hilly areas, the large topographical undulations result in sparse sensor deployment and severely incomplete historical data. At the same time, the region experiences large interannual climate fluctuations and unstable flowering periods, making manual data collection difficult. This leads to significant prediction errors in traditional models, rendering them unsuitable as a basis for pre-production management and market scheduling.

[0004] Therefore, how to conduct intelligent modeling based on multi-source heterogeneous data in the case of missing or abnormal data, and construct a high-precision peach yield prediction method suitable for mountainous planting scenarios, has become an important technical problem that smart orchards urgently need to solve. Summary of the Invention

[0005] The purpose of this invention is to provide a method for predicting peach yield based on comprehensive data analysis, in order to address the shortcomings of the prior art.

[0006] To achieve the above objectives, the present invention provides the following technical solution: a method for predicting peach yield based on comprehensive data analysis, comprising: Obtain the multi-source dataset D for the target orchard area; Based on dataset D, a dynamic climate anomaly index E for orchard areas is constructed to measure the impact of extreme climate events at different times of the year on the fruit tree development cycle in the target area. The extracted fruit tree physiological response data were converted into a set of micro-region physiological state response parameters, and combined with the index E, a bimodal graph structure G for the orchard area was established. In the graph, one mode represents the topographic relationship and the other mode represents the fruit tree physiological response relationship. The bimodal graph G is input into the graph neural network model F, and the spatiotemporal features between nodes in the region are extracted using F to obtain the feature tensor T; The overall information entropy change rate ΔH of the orchard is calculated based on the feature tensor T. If ΔH exceeds the set threshold, the edge weight matrix and the number of propagation layers in the graph neural network model F are dynamically adjusted to obtain the adaptive model F′. The prediction results output by model F′ are grouped according to the spatial regions of the orchard, and the predicted yield of each sub-region is calculated separately. Finally, the results are combined to obtain the total orchard yield P.

[0007] Preferably, the dataset D includes historical meteorological data, fruit tree microenvironment data, remote sensing image data, soil electrical conductivity data, and fruit tree physiological response data, wherein the fruit tree physiological response data includes trunk sap flow density, leaf transpiration rate, and flowering hormone levels.

[0008] Preferably, the dynamic climate anomaly index E for the orchard area includes: Based on historical meteorological data in dataset D, the long-term climate mean μ and standard deviation σ of the target orchard area on various meteorological parameters are calculated according to the yearly and daily time scales, and a climate benchmark reference system is constructed. The daily meteorological data for the target year are compared with the corresponding long-term mean μ, and the standardized climate deviation value is calculated using the Z-score method to measure the degree of meteorological anomaly on that day. Setting thresholds for judging extreme weather events For any meteorological parameter, when the absolute value of its Z-value is greater than 1, When an event occurs, it is considered a local extreme event, and weights are assigned according to its duration, intensity, and stage of occurrence to construct its corresponding abnormal event factor. By fusing anomalous event factors from multiple parameters, including temperature, precipitation, humidity, and sunshine, a comprehensive climate anomaly index E is constructed using a weighted normalized aggregation function.

[0009] Preferably, the establishment of the bimodal graph structure G for the orchard area includes: The collected sap flow density, leaf transpiration rate, and flowering hormone levels were clustered according to spatial location to form micro-regional physiological state units with geographic labels. For each micro-region physiological state unit, the mean, rate of change and fluctuation amplitude of its physiological parameters within a fixed time window are calculated to form the micro-region physiological state response parameter set R. The parameter set R is associated with the dynamic climate anomaly index E. Based on the similarity between orchard topology and physiological state, spatial adjacency graph and state similarity graph are constructed respectively. By integrating spatial adjacency graphs and state similarity graphs, and employing graph overlay and weight normalization strategies, a bimodal graph structure G for the orchard region is established.

[0010] Preferably, the fusion of spatial adjacency graph and state similarity graph includes: Using fruit trees or micro-region units as nodes, construct a spatial adjacency graph based on their geographical coordinates. The edge weights are calculated based on the reciprocal of the Euclidean distance; Based on the physiological state response parameter set R of each microregion, cosine similarity or Pearson correlation coefficient between nodes is calculated to construct a state similarity graph. The edge weight represents the similarity of physiological characteristics; For the picture and A weighted fusion process is performed, and a linear superposition strategy is used to generate the fused graph G′. The edge weights of the fused graph G′ are normalized to obtain the final bimodal graph structure G.

[0011] Preferably, the bimodal graph G is input into the graph neural network model F, including: Using the micro-region physiological state response parameter set R of each node in the bimodal graph G as the initial feature vector, and combining it with the historical time series data of the node, the temporal feature matrix X of the graph node is constructed. The graph structure G and the feature matrix X are input into the graph neural network model F. The model F uses a graph convolutional network or a graph attention network as a spatial feature extraction module, and combines a gated recurrent unit or a temporal convolutional network as a temporal dynamic modeling module. In model F, multiple rounds of adjacency propagation and feature aggregation operations are performed on all nodes in the graph to fuse spatial adjacency information and temporal evolution features, generating a high-dimensional spatiotemporal embedding representation. The spatiotemporal embedding representations output by each node are sequentially concatenated to form the feature tensor T.

[0012] Preferably, the adaptive model F′ is obtained by dynamically adjusting the edge weight matrix and the number of propagation layers in the graph neural network model F, including: Based on the temporal stability of node representations in feature tensor T, the rate of change of state information entropy between nodes is calculated. When the rate of change of information entropy exceeds the set threshold θ, update the normalization strategy of the edge weight matrix in the graph structure. The number of propagation layers L in the graph neural network model F is adaptively adjusted based on the overall graph state complexity index C. The state complexity C is the aggregate value of the embedding dissimilarity of all nodes in the graph. The updated edge weight matrix and propagation layer configuration are used as inputs to reconstruct the graph neural network structure, forming an adaptive model F′.

[0013] Preferably, the final total orchard yield P is obtained by merging the following: Based on orchard geographic information system data, the orchard is divided into several spatial sub-regions. Each sub-region contains multiple graph neural network nodes, and there is a unique spatial mapping relationship between the nodes and the region. The predicted values ​​of each node output by model F′ are aggregated to the corresponding sub-region, and the predicted output of each sub-region is calculated using the area-weighted average method. ; Based on the actual managed area of ​​each sub-region With standardized weighting factors Adjusting for forecast bias, we obtain the corrected regional output. ; The total orchard yield P is calculated by weighted summation of the corrected yields of all sub-regions.

[0014] The technical effects and advantages provided by the present invention in the above technical solution are as follows: 1. This invention introduces fruit tree physiological response data and dynamic climate anomaly indices, integrating spatial topology and physiological state similarity to construct a graph neural network input graph with a dual-modal structure. This effectively overcomes the expressive capacity bottleneck of traditional models when processing multi-source heterogeneous agricultural data. Through the adaptive adjustment mechanism of the edge weight matrix and propagation layer number in the graph neural network, the model possesses the ability to dynamically respond to climate changes and regional instability, significantly improving the timeliness and robustness of yield prediction, and is particularly suitable for intelligent orchard management in mountainous and microclimate environments.

[0015] 2. This invention introduces a spatial grouping and area-weighted fusion strategy in the model output stage, realizing the transformation from single-node prediction to regional yield integration, giving the prediction results clear geographical attribution and actionable decision-making value. This method not only improves the accuracy and interpretability of peach yield prediction, but also provides strong data support for orchard precision fertilization, harvest scheduling, and market supply early warning. Attached Figure Description

[0016] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in this invention. For those skilled in the art, other drawings can be obtained based on these drawings.

[0017] Figure 1 This is a flowchart of the method of the present invention. Detailed Implementation

[0018] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, 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, 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.

[0019] For examples, please refer to Figure 1 As shown in this embodiment, a method for predicting peach yield based on comprehensive data analysis includes: Obtain the multi-source dataset D for the target orchard area; Based on dataset D, a dynamic climate anomaly index E for orchard areas is constructed to measure the impact of extreme climate events at different times of the year on the fruit tree development cycle in the target area. The extracted fruit tree physiological response data were converted into a set of micro-region physiological state response parameters, and combined with the index E, a bimodal graph structure G for the orchard area was established. In the graph, one mode represents the topographic relationship and the other mode represents the fruit tree physiological response relationship. The bimodal graph G is input into the graph neural network model F, and the spatiotemporal features between nodes in the region are extracted using F to obtain the feature tensor T; The overall information entropy change rate ΔH of the orchard is calculated based on the feature tensor T. If ΔH exceeds the set threshold, the edge weight matrix and the number of propagation layers in the graph neural network model F are dynamically adjusted to obtain the adaptive model F′. The prediction results output by model F′ are grouped according to the spatial regions of the orchard, and the predicted yield of each sub-region is calculated separately. Finally, the results are combined to obtain the total orchard yield P.

[0020] In this invention, the first step is to acquire a multi-source dataset D of the target orchard area to construct the subsequent yield prediction model. This dataset D comprehensively considers environmental factors, crop physiological responses, and remote sensing imagery information, specifically including the following five key data categories: Historical meteorological data includes daily meteorological observations from at least the past 3 to 5 years. Data fields include, but are not limited to: daily average temperature, extreme maximum / minimum temperature; daily precipitation; relative humidity; sunshine duration and solar radiation intensity; wind speed and direction. The above data can be collected by meteorological monitoring stations deployed near the orchard, obtaining spatially interpolated gridded meteorological data with a temporal resolution of no less than daily and a spatial resolution of no less than 0.1 degrees.

[0021] Environmental monitoring nodes are deployed within the orchard to form a microclimate sensing network for collecting the following data: orchard canopy temperature and humidity; surface and soil temperature; photosynthetically active radiation intensity; and carbon dioxide concentration. This microenvironment data acquisition equipment can utilize commercial agricultural sensor stations, such as LoRa / 4G IoT sensors from brands like HOBO, Aranet, and Ruinong. The distance between each monitoring node should not exceed 30 meters, and the data collection frequency is once every 10 minutes.

[0022] Use drones or satellite remote sensing platforms (such as Sentinel-2 and PlanetScope) to regularly acquire remote sensing image data of the orchard area. Images should include: visible light (RGB); multispectral data (NDVI, EVI, and other vegetation indices); thermal infrared data (for assessing tree water stress); and three-dimensional digital terrain models (DTM). It is recommended that remote sensing data be acquired every 7 days, with time window registration performed in conjunction with key fruit tree growth stages (budding, flowering, fruit enlargement, and ripening) to create a time-series image sequence.

[0023] Obtain soil conductivity profiles in orchards using handheld or buried soil conductivity meters (such as Veris or EM38). The recommended sampling depth is 0–60 cm, with stratified sampling (e.g., 0–20 cm, 20–40 cm, 40–60 cm). Collect at least 5 samples per acre and construct spatial interpolation maps based on geographical location to assess soil moisture and salinity heterogeneity.

[0024] The physiological response data of fruit trees described in this invention mainly reflects the true physiological adaptation state of fruit trees to environmental factors, including the following three categories: Sap Flow Density: Sap flow sensors were installed on the trunks of representative fruit trees using a thermal diffusion probe method to monitor the water transport rate in real time, with a frequency of 15 minutes per measurement.

[0025] Leaf transpiration rate: Use a portable stomatal meter (such as LI-6400) to measure the rate regularly on representative leaves of different tree ages and regions. It is recommended to monitor once every 7 to 10 days and collect no less than 50 samples per orchard.

[0026] Hormone levels during flowering: Samples of newly formed flower buds were collected at the initial flowering stage and the full bloom stage. The contents of gibberellin (GA), auxin (IAA), ethylene (ETH), etc. were detected by high performance liquid chromatography-mass spectrometry (HPLC-MS). Fluctuations in hormone levels were extracted as early warning signals for abnormal flowering.

[0027] All of the above data needs to be standardized in format and synchronized in time, and uniformly organized into a structured dataset D, with fields aligned by timestamps and spatial coordinates bound to the orchard GIS system.

[0028] In this invention, a dynamic climate anomaly index E for orchard areas is constructed based on dataset D to measure the impact of extreme climate events at different times of the year on the fruit tree development cycle in the target area.

[0029] Extract daily meteorological data for at least five consecutive years from the target orchard area from the multi-source dataset D. Data fields include daily average temperature, daily precipitation, relative humidity, and sunshine duration.

[0030] Using time as the horizontal axis, the multi-year average value μ and standard deviation σ of each type of meteorological factor are calculated based on historical data samples from the "same date", thereby establishing a daily-granular climate benchmark reference system.

[0031] In the target year, daily meteorological observation data are acquired in real time from the orchard's current meteorological sensor network. The actual daily observation values ​​for each day of the year are compared with the corresponding long-term baseline mean, and the standardized deviation Z of each indicator is calculated using the standard deviation normalization method (Z-score method). The formula for calculating the Z-score is: Where X is the measured meteorological value on a specific day in the current year, and μ and σ are the historical mean and standard deviation for that day, respectively. The Z-value measures the statistically significant degree of anomalousness of the climate on that day. The larger the absolute value of the Z-value, the more the current climate deviates from the normal level.

[0032] To determine whether an extreme weather event has occurred, this invention sets an anomaly identification threshold. It is recommended to set the default value to 2.0, meaning that when the Z value on a certain day is greater than or equal to ±2.0, it is considered a statistical outlier. When judging extreme events, the time windows in which Z value anomalies occur consecutively are aggregated, and the following three indicators are defined: Intensity factor: equal to the absolute value of the Z value; Duration factor: the number of consecutive days the anomaly occurs; Stage factor: whether the event occurs during key phenological periods such as flowering period and swelling period.

[0033] Each extreme weather event will be assigned an event score based on the three factors mentioned above. A higher score indicates a potentially greater impact on yield. The event score can be calculated using the following weighted average: Event Score = α × Intensity + β × Duration Days + γ × Stage Weight; where α, β, and γ are empirical coefficients (e.g., α=1, β=0.5, γ=2), and the stage weight is set to 2-3 during critical phenological periods and 1 during non-critical periods.

[0034] To comprehensively reflect the overall situation of climate anomalies encountered by the orchard throughout the year, the scores of all extreme events throughout the year are accumulated, and the importance weights of different meteorological factors are introduced to construct the total index E.

[0035] Suppose that m extreme events are identified throughout the year, and the score for each event is... Then we can define: Total Abnormal Impact Score , i = 1 to m.

[0036] Then, the meteorological factors (e.g., temperature 40%, precipitation 30%, humidity 15%, sunshine 15%) are incorporated into the weighted model to calculate the comprehensive dynamic climate anomaly index E, as shown in the formula: Where S represents the cumulative score for different weather types, and W represents the corresponding weight. The index E generally ranges from 0 to 100; the larger the value, the more severe the climate anomaly experienced by the orchard.

[0037] In one embodiment of the present invention, in order to more comprehensively characterize the spatial positional relationship and physiological state response characteristics of each fruit tree unit in the orchard area, a dual-modal graph structure G is proposed to be established. This graph structure integrates the topographic relationship of the terrain and the physiological response relationship of the fruit trees, and serves as the input basis for subsequent graph neural network modeling.

[0038] First, based on the physiological response data of fruit trees obtained in the aforementioned steps, including: trunk sap flow density (reflecting water transport capacity); leaf transpiration rate (reflecting water loss and stomatal opening); and flowering hormone levels (such as gibberellin and ethylene, used to identify the activity status during flowering), the collection points are clustered according to spatial coordinates. For example, density-based spatial clustering algorithms (DBSCAN) or K-means algorithms are used to divide the fruit trees in the orchard into several micro-regions. Each micro-region contains a group of fruit trees that are spatially close and in similar states, and each micro-region is assigned a unique geographical label.

[0039] For each micro-region physiological state unit, within a set time window (such as 7 or 15 consecutive days), the physiological response data of each fruit tree within it are statistically analyzed to extract the following three types of features: mean: representing the central trend of the current physiological state; rate of change: for example, the daily change rate of sap flow density, calculated by averaging the daily differences to reflect the rate of change of state; fluctuation range: representing the stability of the physiological response within the time window, such as the difference between the maximum and minimum values, standard deviation, etc.

[0040] Using the above methods, a structured parameter set R can be generated for each microregion to describe the physiological response characteristics of that region within a specific time period.

[0041] After obtaining the parameter set R for all microregions, we further construct graph structures for two different modes: Constructing a spatial adjacency graph Treating each micro-region as a node in the graph, calculate the Euclidean distance between any two nodes based on the geographical coordinates of their center points. Let the spatial distance between two nodes i and j be d(i, j), then their edge weights... Defined as: ,when An edge is established when the adjacency threshold T is less than the set threshold value; otherwise, no edge is established. This graph structure... Used to describe the spatial proximity between fruit trees, emphasizing topographical or locational relevance.

[0042] Construct a state similarity graph Using the parameter set R of each microregion as a vector, the physiological state similarity between any two nodes can be calculated. One of the following two commonly used metrics can be chosen: cosine similarity or Pearson correlation coefficient. Taking cosine similarity as an example, let the state parameter vectors of microregions i and j be... and Then its border rights Defined as: Where "·" represents the vector dot product, and "||" represents the vector magnitude. Larger edge weights indicate more consistent physiological responses between the two regions. The resulting graph structure... Used to capture the synchronous response of fruit trees to external climate or internal moisture conditions.

[0043] To simultaneously consider spatial topology and physiological state response, this invention proposes to... and Perform graph fusion. This involves the following two steps: For any two nodes i and j, and The corresponding edge weights are respectively and Define the fusion of edge weights for: Where s1 and s2 are weighting coefficients, and s1 + s2 = 1. These coefficients can be dynamically set according to actual climate anomalies. For example, when the climate anomaly index E is high, the value of s2 is set larger to increase the weight of physiological responses; when E is low, the weight of spatial relationships s1 can be appropriately increased. For example, when E is between 0 and 100, the following setting is used: .

[0044] To ensure the numerical stability of the graph structure in subsequent graph neural networks, all edge weights in the fused graph G′ are subjected to max-min normalization, mapping them to the range of 0 to 1, resulting in the final bimodal graph structure G. The normalization formula is: Where w_min and w_max are the minimum and maximum values ​​of all edge weights in graph G′, respectively. The resulting graph structure G includes both the spatial proximity between fruit trees in the orchard and the synchronous changes in the physiological states of each region, possessing strong expressive power, and serves as the input structure for subsequent graph neural network modeling.

[0045] Each node in the constructed bimodal graph structure G is considered a micro-region physiological state unit. For each node, multidimensional physiological response data within a past time window (e.g., 7 or 14 consecutive days) are extracted, including: sap flow density; leaf transpiration rate; and flowering hormone levels. These parameters are arranged in chronological order to construct a node feature time series. Assuming each node contains D-dimensional features and the time window is T, the size of the temporal feature matrix constituting each node is T×D. The features of all nodes are concatenated to form the input feature matrix X, with dimensions N × T × D, where N is the total number of nodes in the graph.

[0046] The graph neural network model F includes two types of sub-modules: Spatial Feature Extraction Module: This module employs either a Graph Convolutional Network (GCN) or a Graph Attention Network (GAT) to process the spatial topological information in the graph structure G. If GCN is used, its core operation is to perform a weighted average of the features of each node with the features of its neighboring nodes. The calculation formula is: Output of each layer = Activation Function (Graph Normalized Adjacency Matrix × Input Feature Matrix × Weight Matrix); where the graph normalized adjacency matrix is ​​calculated based on the fused edge weights in the bimodal graph G, and the weight matrix is ​​a learnable parameter. If GAT is used, an attention mechanism is introduced, dynamically calculating attention coefficients based on the feature correlations between nodes, and then weighted aggregating the features of neighboring nodes, effectively improving the expressive power for heterogeneous graph structures.

[0047] Temporal dynamic modeling module: Considering the evolutionary trend of fruit tree physiological state over time, this invention introduces a temporal modeling module after spatial graph convolution. Two preferred structures are: Gated Recurrent Unit (GRU) for capturing long- and short-term temporal dependencies; or Temporal Convolutional Network (TCN) for parallel modeling of feature trends at different time steps. The model inputs the feature vector of each node into the GRU or TCN unit, encodes the state at T time steps, and outputs the temporal embedding features of each node.

[0048] The graph neural network model F includes two types of sub-modules: Spatial Feature Extraction Module: This module employs either a Graph Convolutional Network (GCN) or a Graph Attention Network (GAT) to process the spatial topological information in the graph structure G. If GCN is used, its core operation is to perform a weighted average of the features of each node with the features of its neighboring nodes. The calculation formula is: Output of each layer = Activation Function (Graph Normalized Adjacency Matrix × Input Feature Matrix × Weight Matrix); where the graph normalized adjacency matrix is ​​calculated based on the fused edge weights in the bimodal graph G, and the weight matrix is ​​a learnable parameter. If GAT is used, an attention mechanism is introduced, dynamically calculating attention coefficients based on the feature correlations between nodes, and then weighted aggregating the features of neighboring nodes, effectively improving the expressive power for heterogeneous graph structures.

[0049] Temporal dynamic modeling module: Considering the evolutionary trend of fruit tree physiological state over time, this invention introduces a temporal modeling module after spatial graph convolution. Two preferred structures are: Gated Recurrent Unit (GRU) for capturing long- and short-term temporal dependencies; or Temporal Convolutional Network (TCN) for parallel modeling of feature trends at different time steps. The model inputs the feature vector of each node into the GRU or TCN unit, encodes the state at T time steps, and outputs the temporal embedding features of each node.

[0050] In this invention, to further improve the modeling ability of graph neural network model F in handling dynamic changes and non-stationary inputs in orchard areas, an adaptive model structure mechanism based on feature fluctuation is proposed. By dynamically adjusting the edge weight matrix and the number of propagation layers of the model, an adaptive model F′ with time-varying response capability is generated.

[0051] First, based on the feature tensor T output from the previous stage, the state representation of each node in the graph is extracted at multiple time steps. Assume there are N nodes in the graph, the time window length is T, and the feature representation of each node at each time point is a vector.

[0052] For each node, its feature distribution within the two most recent time windows is statistically analyzed, and the state uncertainty at each moment is calculated using Shannon information entropy. Then, the rate of change of information entropy ΔH for that node is obtained by calculating the entropy difference between the two time windows, as shown in the following formula: Information Entropy rate of change of entropy ;in, This represents the distribution probability of each dimension after the node feature vector is normalized. When the average ΔH of a node or the whole exceeds a set threshold θ (e.g., 0.3), it indicates that there is significant state instability or external disturbance in the graph, and the model structure adaptation mechanism should be triggered.

[0053] After triggering the structural adjustment mechanism, the edge weight matrix in graph G is first adaptively reconstructed. The original edge weight matrix is ​​usually processed using a symmetric normalization method, that is, the normalized matrix A′ is calculated using the degree matrix D and the adjacency matrix A, as follows: .

[0054] This invention introduces a weighted reinforcement normalization method, which readjusts the weights of a node's neighbors based on the node's entropy change rate ΔH, as shown in the formula: ;in, Let λ be the original edge weight, λ be the adjustment sensitivity coefficient (ranging from 0.5 to 1.5), and ΔH(i) be the rate of change of information entropy of node i. This method strengthens the influence of adjacent edges of nodes with drastic state changes, thereby improving the model's ability to focus on dynamic regions.

[0055] To balance modeling depth with the risk of overfitting, this invention introduces a graph state complexity metric C to guide the dynamic setting of the number of propagation layers L in the graph neural network. State complexity C is defined as the average degree of difference between the feature representations of all nodes in the graph, and is calculated using the following formula: For all node pairs (i, j), that is, in the feature tensor T, the larger the mean Euclidean distance between any two node representations, the more dispersed the state distribution and the higher the complexity.

[0056] Based on the value of complexity C, the number of propagation layers L is adaptively set. For example, if C < 0.5, it indicates that the states are concentrated, so L = 2 can be set; if 0.5 ≤ C < 1.0, L = 3 can be set; if C ≥ 1.0, it indicates that the states are highly dispersed, so L = 4 can be set. This strategy can avoid overfitting the model when the states are simple or underfitting when the states are complex.

[0057] Using the updated edge weight matrix and the number of propagation layers as input parameters, the graph neural network model structure is reconstructed, resulting in the adaptive model F′. Model F′ maintains the modular architecture of the original model F, adjusting only the following two parts: the adjacency matrix input of the spatial graph convolution module is changed to a weighted normalized matrix; and the number of network layers is reconfigured according to the current state complexity C. Model F′ will be used for subsequent yield prediction and regional attribution analysis tasks, and can be continuously updated over a rolling time window to achieve continuous modeling and response to the dynamic state of the orchard.

[0058] In this invention, in order to fully utilize the prediction results output by the graph neural network model F′ at the micro-region granularity and further improve the spatial expressiveness and manageability of the yield prediction results, the specific implementation steps are as follows: Based on existing Geographic Information System (GIS) data of the orchard, the overall orchard area is divided into several spatial sub-regions with management logic or geographical consistency. Regional division can be based on one or a combination of the following rules: topographic distribution (high and low slopes, drainage direction); tree age structure or variety distribution; actual management units (such as zoned irrigation areas, work group responsibility areas); and unified cultivation history. Each spatial sub-region has a unique region number and boundary coordinate information in the GIS. Simultaneously, the graph nodes (i.e., micro-region units) in the graph neural network model F′ are mapped to their corresponding sub-regions according to their spatial center point coordinates, establishing a one-to-one association between nodes and regions.

[0059] In the preceding steps, model F′ has already output the corresponding predicted yield value for each graph node, denoted as . Let represent the predicted output of the i-th node (units can be kilograms, jin, etc.). For each sub-region, the predicted values ​​of all nodes belonging to that region are aggregated. Since the area of ​​the fruit tree group represented by each graph node may be different, an area-weighted average is used to calculate the total predicted output of the region during aggregation. The formula is as follows: ;in, The predicted value for the i-th node is... The area covered by the fruit trees corresponding to this node (which can be estimated by calculating the crown width from remote sensing images) is represented by Σ, which represents the summation of all nodes in this area.

[0060] Considering the potential for boundary errors, uneven sensor deployment density, or imbalanced training samples in different regions, this invention introduces a region weighting factor. This is used to fine-tune the predicted values ​​for each sub-region. The weighting factor can be set based on the following factors: the actual physical area of ​​the sub-region; historical average annual yield; the percentage of sample nodes; and tree age structure density, etc.

[0061] Regional weight factors are constructed by pre-setting empirical models or machine learning regression methods. Then, the original region yield Make corrections to obtain the final regional output. : The revised version This is more in line with the actual production trends and stability of the region.

[0062] The weighted sum of the revised yield results for all sub-regions yields the overall orchard yield forecast, P. The integrated formula is as follows: ; where Σ represents the summation over all sub-regions. This integration process preserves the weighted distribution of prediction contributions between regions, avoiding the distortion of overall output by extreme prediction points.

[0063] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application.

Claims

1. A method for predicting yield of watermelon based on comprehensive data analysis, characterized in that: The method comprises the following steps: acquiring a multi-source data set D of a target orchard area; constructing a dynamic climate anomaly index E of the orchard area based on the data set D, for measuring the influence of extreme climate events at different time periods within a year on the development cycle of fruit trees in the target area; converting the extracted physiological response data of fruit trees into a micro-area physiological state response parameter group, and combining the index E to establish a dual-mode graph structure G of the orchard area, one mode of the graph being a topological relationship of the terrain and the other mode being a physiological response relationship of the fruit trees; inputting the dual-mode graph G into a graph neural network model F, and extracting the spatio-temporal features between nodes in the area by using the F to obtain a feature tensor T; calculating a change rate ΔH of the overall information entropy of the orchard based on the feature tensor T, and if the ΔH exceeds a set threshold, dynamically adjusting the edge weight matrix and the propagation layer number in the graph neural network model F to obtain an adaptive model F'; grouping the prediction results output by the model F' according to the spatial areas of the orchard, respectively calculating the predicted yield of each sub-area, and finally merging to obtain the overall yield P of the orchard.

2. The method for predicting yield of juicy peach based on comprehensive data analysis according to claim 1, characterized in that: The data set D comprises historical meteorological data, fruit tree micro-environment data, remote sensing image data, soil conductivity data and fruit tree physiological response data, wherein the fruit tree physiological response data comprises sap flow density, leaf transpiration rate and flowering hormone level.

3. The method for predicting yield of juicy peach based on comprehensive data analysis according to claim 1, characterized in that: The construction of the dynamic climate anomaly index E of the orchard area comprises the following steps: based on the historical meteorological data in the data set D, calculating the long-term climate mean μ and the standard deviation σ of each meteorological parameter in the target orchard area according to the year-by-year and day-by-day time scales, and constructing a climate benchmark reference system; comparing each daily meteorological data in the target year with the corresponding long-term mean μ, and calculating the standardized climate deviation value by using the Z-score method to measure the abnormal degree of the daily meteorological data; Setting extreme weather event judgment threshold For any meteorological parameter, when the absolute value of its Z value is greater than , it is considered as a local extreme event, and its duration, intensity and occurrence stage are weighted to construct its corresponding abnormal event factor; fusing the abnormal event factors of multiple parameters such as temperature, precipitation, humidity and sunshine, and constructing a comprehensive climate anomaly index E by using a weighted normalized aggregation function.

4. The method for predicting yield of juicy peach based on comprehensive data analysis according to claim 1, characterized in that: The establishment of the dual-mode graph structure G of the orchard area comprises the following steps: clustering and grouping the collected sap flow density, leaf transpiration rate and flowering hormone level according to the spatial position to form micro-area physiological state units with geographical labels; for each micro-area physiological state unit, calculating the mean, change rate and fluctuation amplitude of the physiological parameters within a fixed time window to form a micro-area physiological state response parameter group R; associating and modeling the parameter group R and the dynamic climate anomaly index E, and based on the topological structure of the orchard and the similarity of the physiological state, constructing a spatial adjacency graph and a state similarity graph respectively; fusing the spatial adjacency graph and the state similarity graph, and establishing the dual-mode graph structure G of the orchard area by using the graph superposition and weight normalization strategy.

5. The method for predicting yield of juicy peach based on comprehensive data analysis according to claim 4, characterized in that: The fusion of the spatial adjacency graph and the state similarity graph comprises the following steps: The fruit trees or micro-area units are taken as nodes, and a spatial adjacency graph is constructed according to geographical coordinates of the nodes The edge weight is calculated according to the reciprocal of Euclidean distance. Based on the physiological state response parameter group R of each micro area, cosine similarity or Pearson correlation coefficient between nodes is calculated to construct a state similarity graph , and the edge weight represents physiological feature similarity. To the Figure And Weighted fusion is performed, and a linear superposition strategy is adopted to generate a fusion image ; to the fusion graph The edge weights of the fusion graph are normalized to obtain the final bimodal graph structure G.

6. The method for predicting yield of juicy peach based on comprehensive data analysis according to claim 5, characterized in that: The inputting of the dual-mode graph G into the graph neural network model F comprises the following steps: taking the micro-area physiological state response parameter group R of each node in the dual-mode graph G as an initial feature vector, and combining the historical time series data of the node to construct a time series feature matrix X of the graph node; inputting the graph structure G and the feature matrix X into the graph neural network model F, wherein the model F uses a graph convolution network or a graph attention network as a spatial feature extraction module, and combines a gated recurrent unit or a time convolution network as a time dynamic modeling module; In model F, multi-round neighbor propagation and feature aggregation operations are performed on all nodes in the graph to fuse spatial neighbor information and time evolution features, generating high-dimensional spatiotemporal embedding representations; The spatiotemporal embedding representations output by each node are sequentially spliced to form a feature tensor T.

7. The method for predicting yield of juicy peach based on comprehensive data analysis according to claim 6, characterized in that: Among them, the edge weight matrix and the number of propagation layers in the dynamic adjustment graph neural network model F are obtained to obtain an adaptive model F', including: Based on the time sequence stability of the node representation in the feature tensor T, the information entropy change rate between nodes is calculated; When the information entropy change rate exceeds the set threshold θ, the normalization strategy of the edge weight matrix in the graph structure is updated; According to the state complexity index C of the whole graph, the number of propagation layers L in the graph neural network model F is adaptively adjusted, and the state complexity C is the aggregation value of the embedding difference of all nodes in the graph; The updated edge weight matrix and the number of propagation layers are configured as inputs to reconstruct the graph neural network structure to form an adaptive model F'.

8. The method for predicting yield of juicy peach based on comprehensive data analysis according to claim 7, characterized in that: Finally, the total yield P of the orchard is obtained, including: Based on the geographic information system data of the orchard, the orchard is divided into several spatial sub-regions, each sub-region contains multiple graph neural network nodes, and the nodes and the regions have a unique spatial mapping relationship; Each node prediction value output by the model F' is aggregated to the corresponding sub-region, and an area weighted average method is used to calculate the predicted yield of each sub-region ; According to the actual management area of each sub-region With the standardized weight factor , adjust the prediction value deviation to obtain the corrected regional yield ; The corrected yield of all sub-regions is weighted and summed to calculate the total yield P of the orchard.

Images