Method for identifying and controlling orchard area boundary based on meteorological terrain and tree age characteristics

Abstract

The present application relates to the field of smart agriculture and precise orchard management technology, especially the method for identifying and controlling the boundary of the orchard region based on meteorological topography and tree age characteristics, and the technical scheme comprises: obtaining the time sequence meteorological data, topographic digital elevation model and remote sensing image of the orchard; using a deep learning visual detection model to extract the crown and trunk characteristics of individual fruit trees, combining meteorological and topographic data to output individual tree age values and spatial distribution maps through a tree age regression model; constructing a meteorological-topographic-tree age coupled digital twin scene, generating a comprehensive feature tensor through multi-modal feature fusion; inputting the tensor into a boundary deduction model based on reinforcement learning or genetic algorithm, taking ecological adaptability and operational economy as double targets, and dynamically outputting multi-level control boundaries of the core growth area, the buffer adjustment area and the mechanical operation restriction area; effectively fusing multi-source heterogeneous information, and significantly improving the precision of orchard partition and the intelligent level of control decision.

Description

Technical Field

[0001] This invention relates to the field of smart agriculture and precision orchard management technology, and in particular to a method for orchard area boundary identification and control based on meteorological topography and tree age characteristics. Background Technology

[0002] Accurately differentiating the production potential and operational suitability of different areas within an orchard, and implementing differentiated water and fertilizer management and mechanized operations, is the core development direction of current intelligent orchard management. Currently, the technologies related to orchard zoning and control mainly fall into the following categories: The first category is static zoning methods based on a single topographic or soil attribute. Existing technologies typically use digital elevation models to extract topographic factors such as slope and aspect, or perform spatial interpolation using physicochemical properties such as soil electrical conductivity and organic matter content, thereby dividing the orchard into several operational units. However, such methods only focus on the site conditions of the land itself, completely ignoring the decisive influence of individual growth differences in fruit trees as living organisms (especially tree age) on management strategies. Fruit trees of different ages exhibit significant differences in nutrient requirements, root distribution, pruning methods, and yield composition. Simply relying on topographic or soil zoning cannot achieve the precise management goal of "tailoring policies to individual trees."

[0003] The second category is fruit tree growth monitoring and age estimation technology based on remote sensing imagery. In recent years, some studies have attempted to extract phenotypic parameters such as crown width and color index of fruit trees using UAV or satellite remote sensing imagery and establish empirical regression models with tree age. However, existing tree age estimation methods have significant shortcomings: on the one hand, the crown width-age growth relationship is strongly affected by local topography (such as light differences caused by slope aspect and temperature gradients caused by altitude) and interannual meteorological fluctuations (such as growth suppression in drought years). Traditional statistical regression models have failed to effectively incorporate these spatial heterogeneous factors, resulting in a serious lack of generalization ability of tree age estimation under cross-regional and cross-topographical conditions; on the other hand, existing methods mostly remain at the level of digital extraction of single tree age, lacking a systematic technical solution for deeply integrating it with topographic and meteorological data and applying it to the dynamic delineation of management boundaries.

[0004] The third category is orchard digital twin and virtual simulation technology. Existing orchard visualization methods based on 3D modeling can map sensor data onto virtual scenes for display. However, these digital twin systems are mostly static displays or unidirectional data-driven, lacking mechanism-data joint modeling of the coupling relationship between "meteorology-topography-fruit tree growth," and even more so, lacking the ability to dynamically extrapolate control boundaries based on multi-objective optimization algorithms. Existing twin systems cannot answer the core management question of "under the current weather forecast and tree age distribution, which areas should be prioritized for investment, and which areas should have mechanical operations restricted."

[0005] The fourth category concerns methods for delineating orchard work areas. Existing technologies often rely on fixed grids, roads and ditches, or simple slope thresholds to delineate mechanical access and manual operation areas, and these boundaries remain unchanged for a long period once defined. However, as fruit trees age, their canopy expansion and root competition evolve continuously; simultaneously, meteorological conditions in different years (such as wet and dry years) significantly affect soil carrying capacity and the safety of mechanical access. Existing static boundary delineation methods lack a dynamic optimization mechanism based on the synergistic goals of ecological adaptability and operational economy, failing to adaptively adjust to the fruit tree growth cycle and interannual climate changes, resulting in management strategies lagging behind actual orchard conditions. Furthermore, existing technologies generally lack a complete chain from boundary identification to the generation of differentiated management instructions and then to closed-loop feedback on execution effects, making it difficult to continuously improve the accuracy of boundary delineation and the effectiveness of management decisions.

[0006] In summary, the existing technologies for orchard area boundary identification generally have the following technical defects: (1) They do not organically couple meteorological dynamic factors, topographic spatial factors and individual tree age characteristics, and lack a comprehensive characterization method that reflects the multidimensional heterogeneity of orchards; (2) The accuracy of tree age estimation is greatly affected by topography and meteorology, and a dynamically updatable tree age spatial distribution map has not been constructed; (3) The area boundary is a static and rigid division, and it is impossible to dynamically extrapolate and adaptively update based on the dual objectives of ecology and economy; (4) There is a lack of intelligent mapping method for the entire process from boundary identification to differentiated management and control.

[0007] In view of this, the present invention proposes a method for orchard area boundary identification and management based on meteorological topography and tree age characteristics to solve the existing problems. Summary of the Invention

[0008] The purpose of this invention is to provide a method for orchard area boundary identification and management based on meteorological topography and tree age characteristics, so as to solve the problems mentioned in the background art.

[0009] To achieve the above objectives, the present invention provides the following technical solution: a method for orchard area boundary identification and management based on meteorological topography and tree age characteristics, comprising the following steps: S1: Acquire multi-source heterogeneous data of the target orchard, including time-series meteorological data, topographic digital elevation model data, and orchard remote sensing image data containing individual morphological characteristics of fruit trees; S2: Based on deep learning network, a visual detection model is built to extract the crown and trunk features of each fruit tree from orchard remote sensing image data. Combined with time-series meteorological data and topographic digital elevation model data, the individual tree age value and tree age spatial distribution map of each fruit tree are output through the tree age regression prediction model. S3: Construct a digital twin scene that couples meteorology, topography, and tree age, and perform spatiotemporal registration and feature fusion of time-series meteorological data, topographic digital elevation model data, and tree age spatial distribution map to generate a comprehensive feature tensor that reflects the differences in growth potential and environmental stress in different areas of the orchard. S4: Input the comprehensive feature tensor into the boundary inference model constructed based on reinforcement learning or genetic optimization algorithm, and dynamically infer and output the multi-level dynamic control boundary inside the orchard with ecological adaptability and operational economy as dual objective constraints. The multi-level dynamic control boundary includes the core growth zone, buffer regulation zone and mechanical operation restriction zone. S5: Based on multi-level dynamic control boundaries and combined with real-time or forecast meteorological data, generate differentiated regional control strategies and map the control strategies to the execution instruction set of the orchard operation terminal to realize orchard zoned control; The feature fusion in step S3 employs a multimodal cross-attention fusion network. This network performs cross-attention encoding on temporal meteorological features and topographic features to generate an attention weight map. Then, the attention weight map is summed pixel-by-pixel with preset stress factor base values ​​to generate a comprehensive environmental stress coefficient map. Finally, the environmental stress coefficient map is concatenated with the tree age spatial distribution map along the channel dimension, and then subjected to an adaptive weighted fusion through a 3×3 convolutional layer to generate the final comprehensive feature tensor. The preset stress factor base values ​​include drought stress coefficient, low temperature freezing risk coefficient, and topographic drainage index.

[0010] Furthermore, the time-series meteorological data in step S1 includes historical daily temperature, precipitation, solar radiation, wind speed, and weather forecast data within a future preset time window; the topographic digital elevation model data includes slope, aspect, altitude, and surface roughness factor; and the orchard remote sensing image data includes multispectral imagery, thermal infrared imagery, and visible high-resolution imagery.

[0011] Furthermore, in step S2, the visual detection model adopts the Mask R-CNN instance segmentation network, and the backbone network adopts the Swin Transformer Tiny. On the basis of the backbone network, an attention enhancement module and an edge refinement branch are embedded to extract the crown diameter, crown projection area, trunk diameter and branch distribution density of a single fruit tree. At the same time, the crown change rate is calculated by using the change in crown area of ​​the same single tree detected in two adjacent time series images.

[0012] Furthermore, the tree age regression prediction model is constructed as follows: using the crown width and trunk features extracted by the visual detection model as input, combined with the historical planting year or standard age class sample labels of the corresponding fruit trees, a graph neural network is used for training; and during the training process, the slope aspect and altitude factors in the topographic digital elevation model data are introduced as spatial covariates to eliminate the influence of topographic heterogeneity on the tree age-crown width growth relationship, and the individual tree age value is corrected based on the crown width change rate of adjacent time series images.

[0013] Furthermore, the digital twin scenario includes a physical entity layer, a data mapping layer, and a simulation and inference layer. The physical entity layer is used to access orchard IoT sensor data in real time. The data mapping layer is used to construct a three-dimensional virtual orchard model and overlay a spatial distribution map of tree age. The simulation and inference layer is based on a long short-term memory network or a temporal Transformer model to make rolling predictions on the tree age growth trend and boundary drift under different weather scenarios in the future, and feeds the prediction results back to the boundary inference model to iteratively update the boundary.

[0014] Furthermore, the boundary extrapolation model in step S4 employs a non-dominated sorting genetic algorithm or a deep deterministic strategy gradient algorithm for multi-objective optimization. The quantitative indicators of ecological adaptability include regional light energy interception efficiency, soil moisture availability, and carbon sink accumulation. The quantitative indicators of operational economy include agricultural machinery turning radius loss, fuel consumption rate, and harvesting and transportation distance cost. The multi-level dynamic control boundary is determined based on the optimal compromise solution of the Pareto front solution set.

[0015] Furthermore, the core growth zone is an area where the trees are in their peak fruiting period, the terrain slope is less than the preset threshold, and the environmental stress coefficient is below the lower limit. This area is designated as a priority area for water and fertilizer input and flower and fruit management. The buffer adjustment zone is an area where the trees are young or senescent, or where the environmental stress coefficient is within the preset middle range. This area is designated as a dynamic adjustment agronomic measures area. The mechanical operation restriction zone is an area where the terrain slope is greater than the preset upper limit, or where there is a risk of water accumulation, or where the tree age distribution is uneven. This area is designated as a zone where the volume of machinery entering is restricted and drone operations are adopted.

[0016] Furthermore, the differentiated regional control strategy in step S5 includes: generating variable rate fertilization prescription maps, variable spraying flight path planning maps, and irrigation valve zoning control timing tables based on multi-level dynamic control boundaries; and generating UAV terrain-following flight obstacle avoidance paths and hovering spraying strategies based on boundary constraints for mechanical operation restricted areas.

[0017] Furthermore, it also includes step S6: constructing a closed-loop evaluation module for control effect, which acquires remote sensing images and sensor feedback data of the orchard in the next time series after each control strategy is implemented, calculates the deviation between the actual growth response index and the simulation prediction value in the digital twin scenario; when the deviation exceeds the preset tolerance threshold, it triggers the online incremental learning mechanism of the boundary extrapolation model, uses the deviation data to correct the model parameters, and recalibrates and automatically updates the multi-level dynamic control boundary.

[0018] Compared with the prior art, the beneficial effects of the present invention are: 1. This invention breaks through the limitations of traditional methods that rely solely on single terrain or soil attributes, incorporating time-series dynamic meteorological data, static topographic spatial data, and tree age characteristics reflecting the individual life cycle of fruit trees into a unified framework. A comprehensive feature tensor is generated through a multimodal cross-attention fusion network. This tensor not only contains information about the growth potential of the fruit trees themselves but also couples the spatial distribution of environmental stresses such as drought, low-temperature freezing damage, and water runoff. It can systematically characterize the multidimensional heterogeneity of the "sky-ground-tree" interaction within the orchard, providing a far superior information foundation for subsequent boundary delineation compared to existing static zoning schemes. This effectively solves the technical deficiency of lacking comprehensive representation methods pointed out in the background art, achieving deep fusion and comprehensive representation of multidimensional heterogeneous features of meteorology, topography, and tree age, significantly improving the comprehensiveness and accuracy of orchard area identification.

[0019] 2. This invention employs an instance segmentation network with embedded attention enhancement modules and edge refinement branches in the visual detection stage, enabling accurate extraction of morphological details such as crown width and trunk from high-resolution remote sensing images. More importantly, it innovatively introduces aspect and elevation factors as spatial covariates into the tree age regression prediction model, effectively eliminating the interference of topographic heterogeneity on the crown width-tree age growth relationship. This gives the tree age estimation model high robustness and cross-regional generalization ability under different topographic conditions. Simultaneously, based on a dynamic correction mechanism for the crown width change rate in adjacent time series, it further suppresses the noise impact of interannual meteorological fluctuations on single estimations, resulting in a significantly more accurate spatial distribution map of tree age than traditional empirical regression methods. This greatly improves the estimation accuracy and generalization ability of individual tree ages in orchards with complex topography.

[0020] 3. This invention abandons the static partitioning model of "one-time delineation, long-term immutability" in existing technologies. Instead, it inputs a comprehensive feature tensor into a boundary deduction model based on reinforcement learning or genetic optimization algorithms. It uses ecological indicators such as light energy interception efficiency and soil moisture availability, along with economic indicators such as agricultural machinery fuel consumption, turning losses, and transportation distance, as dual constraints. The optimal compromise solution is dynamically obtained through Pareto multi-objective optimization. This mechanism allows the boundaries of the core growth zone, buffer zone, and mechanical operation restriction zone to automatically drift and redraw with tree age, canopy expansion, and interannual meteorological changes. It consistently ensures that the current boundary delineation results maintain an optimal balance between ecological protection and operational costs, greatly improving the ability of management strategies to track the actual evolution of orchard conditions. It establishes a dynamic boundary deduction mechanism based on the synergy of ecological and economic dual objectives, achieving adaptive updates of management boundaries as fruit tree growth and meteorological changes occur.

[0021] 4. The three-layer digital twin scenario (physical entity layer, data mapping layer, and simulation inference layer) constructed in this invention not only realizes the three-dimensional virtualization and real-time data mapping of the orchard, but also relies on long short-term memory networks or temporal Transformer models to perform rolling predictions of tree age growth trends and boundary drift under multiple meteorological scenarios. The prediction results are then fed forward to the boundary inference model, enabling the management strategy to have the ability to adapt to future extreme weather events such as drought and low temperatures. This design elevates digital twins from the traditional "static visualization display" to the level of "dynamic simulation and forward-looking decision-making," significantly enhancing the resilience and decision-making foresight of the orchard management system in the context of climate fluctuations. It constructs a fully closed-loop digital twin system of "perception-fusion-inference-execution-feedback," enhancing the system's ability to proactively respond to uncertain meteorological scenarios.

[0022] 5. This invention directly transforms the derived multi-level control boundaries into specific executable operational instructions such as variable-rate fertilization prescription maps, variable-rate spraying routes, irrigation valve zoning timing tables, and UAV terrain-following obstacle avoidance paths. This completes the final link in the "decision-execution" process, solving the problem of the lack of a complete execution chain in existing technologies. Particularly for areas with restricted mechanical operations, it generates UAV hovering precision spraying strategies through boundary constraints, providing a practical alternative for mechanized operations in complex terrain areas. More importantly, the closed-loop evaluation and incremental learning mechanism constructed in step S6 can automatically trigger online model correction and boundary recalibration based on the deviation between actual growth response and simulation prediction. This allows the entire system to continuously evolve and improve performance over time, completely changing the shortcomings of traditional methods that "only execute without evaluation, only output without optimization." It achieves precise mapping from macro-regional boundaries to micro-level agricultural machinery terminal instructions and possesses closed-loop self-optimization capabilities.

[0023] 6. This invention establishes a logically clear three-tiered zoning management system based on the growth stage characteristics of different tree ages (young, peak fruiting, and senescent stages), combined with terrain slope and environmental stress coefficients: the core growth zone focuses on precise water and fertilizer supply to high-value fruit trees in their peak fruiting stage; the buffer zone implements dynamic agronomic adjustments for young or senescent fruit trees; and the machinery operation restriction zone strictly avoids high-slope, waterlogged, and high-risk areas, and restricts the entry of large machinery. This zoning design conforms to the biological laws of fruit tree cultivation and fully considers the safety and economy of agricultural machinery operations, possessing strong practical application and promotion value. The zoning rules are clear and agronomically sound, balancing the life cycle of fruit trees with terrain-based operational safety. Attached Figure Description

[0024] Figure 1 This is a schematic diagram illustrating the workflow of the orchard area boundary identification and control method based on meteorological topography and tree age characteristics according to the present invention. Detailed Implementation

[0025] The technical solution of the present invention will be further described below with reference to the accompanying drawings and specific embodiments. Example 1

[0026] like Figure 1 As shown, the orchard area boundary identification and management method based on meteorological, topographical and tree age characteristics includes data acquisition steps, deep learning-based visual detection and tree age estimation of individual fruit trees, construction of a digital twin scene coupled with meteorology, topography and tree age, multi-level dynamic management boundary deduction steps, generation of differentiated regional management strategies and instruction mapping steps, and closed-loop evaluation of management effect and online incremental learning steps.

[0027] In the data acquisition step, the system of this invention is deployed in a large-scale citrus orchard (approximately 500 acres, with hilly terrain, an altitude between 80 and 260 meters, and a slope range of 0° to 35°). The data acquisition specifically includes: (1) Acquisition of time-series meteorological data: Small automatic weather stations (model HOBO U30 or Davis Vantage Pro2) were set up in the central area and surrounding areas of the orchard to continuously record daily temperature (°C), precipitation (mm), and solar radiation (W / m²) at an hourly rate. 2 The system collects wind speed (m / s) and air humidity. Simultaneously, it accesses the API interface of the National Meteorological Administration or a third-party commercial meteorological service via a wireless communication module (4G / 5G or LoRa) to obtain weather forecast data for the next 72 hours and the next 15 days (including hourly precipitation and temperature forecasts). Historical meteorological data must be backed up for at least three complete growth cycles (i.e., three years) to support subsequent temporal feature extraction.

[0028] (2) Acquisition of Digital Elevation Model (DEM) Data: A drone equipped with a LiDAR (such as a DJI M300 RTK equipped with Zenmuse L1) was used to conduct low-altitude terrain-following aerial surveys of the orchard. The flight altitude was 80-120m above the ground, and the flight path overlap rate was set to 60% for the lateral direction and 80% for the heading direction. After acquiring point cloud data, fine DEM data with a spatial resolution of 0.1m was generated by post-processing software (such as DJI Terra or CloudCompare). Furthermore, four terrain factors, namely slope, aspect, elevation, and surface roughness, were extracted by GIS spatial analysis (such as the slope and aspect tools of ArcGIS or QGIS) to form a terrain feature raster layer. The raster size was uniformly resampled to 0.5m×0.5m.

[0029] (3) Orchard Remote Sensing Image Data Acquisition: During the fruit tree growing season (e.g., from the sprouting of spring shoots to the maturation of autumn shoots in citrus trees, i.e., from April to October each year), remote sensing data is collected every 15 days using a drone platform (e.g., DJI Phantom 4 Multispectral) equipped with a multispectral camera (e.g., MicaSense RedEdge-MX) and a thermal infrared camera (e.g., FLIR Vue Pro R). The flight altitude is set at 50m above the ground to ensure a ground resolution (GSD) better than 3cm / pixel. The acquired data includes: multispectral images (blue, green, red, red-edge, and near-infrared bands), thermal infrared images (7.5~13.5μm band, used for canopy temperature inversion), and high-resolution visible light RGB images. After radiometric calibration and geometric correction, all image data are orthophoto stitched using Pix4Dfields or Agisoft Metashape software to generate an orchard orthophoto map (DOM) and digital surface model (DSM) with consistent spatial resolution.

[0030] In the deep learning-based visual detection and age estimation of individual fruit trees, the visual detection model is first constructed and individual tree features are extracted, and then the individual tree age value is regressed and predicted.

[0031] In the construction of the visual detection model and the extraction of single-plant features, the Mask R-CNN instance segmentation network is used as the basic architecture of the visual detection model. The specific implementation method is as follows: The backbone network employs Swin Transformer Tiny (Swin-T) to fully leverage its strong representational ability of the irregular boundaries of the fruit tree canopy through its global self-attention mechanism. Following the backbone network, an Attention Enhancement Module (AEM) is embedded. This module consists of channel attention and spatial attention connected in series, used to strengthen the response weights for the canopy edges and trunk regions. Simultaneously, an Edge Refinement Branch (ERB) is added next to the Mask Head branch. This branch uses an active contour evolution method based on the Deep Snake algorithm to fine-tune the boundaries of the initial mask for instance segmentation, accurately fitting the irregular contours of the fruit tree canopy and reducing undersegmentation or oversegmentation caused by overlapping branches and leaves.

[0032] For training dataset construction: 800 citrus trees were randomly selected from the orchard and neighboring similar orchards and manually labeled. The labels included: the canopy polygon outline, the measured diameter of the trunk base (measured with calipers at 20cm above ground), and the planting year (obtained from orchard planting records). The labeled data was divided into training, validation, and test sets in a 7:2:1 ratio. During training, the AdamW optimizer was used with an initial learning rate of 1e-4, a batch size of 8, and 200 training epochs. Multi-scale training and random rotation data augmentation strategies were employed.

[0033] After training, the high-resolution RGB orthophotos acquired in S1 are input into the visual detection model, which outputs the following for each fruit tree: crown diameter (average of measurements in the east-west and north-south directions), canopy projection area, trunk diameter, and branch density (obtained statistically from edge response intensity within the masked area). Simultaneously, the rate of change of crown area (ΔCA / Δt) is calculated using the change in crown area of ​​the same tree detected in two adjacent time-series images (15 days apart), where ΔCA represents the change in crown area and Δt represents the change over time, used for subsequent tree age correction.

[0034] In individual tree age regression prediction, a graph neural network (GNN) is used to construct the tree age regression prediction model to better capture the spatial correlation between adjacent fruit trees caused by competition or mutual assistance. The specific construction method is as follows: Each fruit tree is considered a node in the graph. The feature vector of a node includes: crown diameter, crown projected area, trunk diameter, branch density, crown change rate extracted by the visual detection model, and topographic factors corresponding to the location of the fruit tree (slope, aspect, and elevation values ​​extracted from the DEM layer). Edges in the graph are constructed based on spatial distance: when the distance between two fruit trees is less than a preset threshold (e.g., 5m), a connection edge is established, and the edge weight is the reciprocal of the distance.

[0035] The training labels for the model are the actual tree age converted from the historical planting year (or, for trees where precise planting records are unavailable, age-class labels are obtained by core sampling of some samples using the growth cone method, with age classes divided into: 1-3 years (juvenile stage), 4-8 years (early fruiting stage), 9-15 years (mid-fruiting stage), 16-25 years (late fruiting stage), and 25 years and above (senescence stage), with the median of each age class assigned as a continuous tree age value). During training, slope aspect and altitude factors from the topographic digital elevation model data are introduced as spatial covariates. Specifically, in the message passing stage of the graph neural network, the slope aspect and altitude features of each node are concatenated as fixed environmental covariates into the feature vector output of each convolutional layer. This allows the model to learn prior topographic-growth relationships such as "for the same crown size, the tree age of fruit trees on sunny slopes is usually smaller than that on shady slopes," thereby eliminating the influence of topographic heterogeneity on the age-crown-growth relationship. The model is trained using the mean squared error loss function and the Adam optimizer.

[0036] After the model is trained and validated, all detected individual fruit trees in the orchard are used as input to output the estimated age of each fruit tree. Then, a spatial distribution map of tree age (raster map, spatial resolution 0.5m×0.5m) covering the entire orchard is generated by Kriging spatial interpolation. The value of each raster represents the estimated age value at that location.

[0037] In the process of constructing a digital twin scenario that couples weather, terrain, and tree age, a three-layer architecture for the digital twin scenario is built: (1) Physical Entity Layer: Soil temperature and humidity sensors (such as TEROS 12, buried at depths of 10cm, 30cm and 60cm), stem flow meters (used to monitor the transpiration rate of fruit trees) and small weather stations are deployed in the orchard. All sensors are aggregated to the edge gateway (such as Huawei AR502H) via RS485 bus, and then the data is uploaded to the cloud digital twin server in real time via 5G communication module. The data refresh frequency is 5 minutes / time.

[0038] (2) Data Mapping Layer: In the Unity 3D or Unreal Engine 3D rendering engine, a 3D virtual terrain model of the orchard is constructed based on the DEM data and DOM images obtained by S1; on the 3D terrain, the spatial distribution map of tree age output by S2 is superimposed and displayed in the form of a pseudo-color heat map. Each fruit tree is instantiated and rendered using a 3D tree model (the model scaling ratio is dynamically adjusted according to its crown width and trunk diameter parameters) to realize the digital twin mapping of the real scene of the orchard.

[0039] (3) Simulation and Prediction Layer: This layer is the core intelligent engine of the digital twin scenario. In this embodiment, the Time Series Transformer model is used as the simulation and prediction model. The model input is the daily time series meteorological data (temperature, precipitation, radiation) for the past 3 years and the current spatial distribution map of tree age. The output is the prediction map of tree age growth trend for the next year (monthly) and the corresponding environmental stress evolution map. Specifically, the Time Series Transformer adopts an encoder-decoder structure. The encoder performs self-attention encoding on the historical meteorological sequence, and the decoder integrates the high-dimensional features of the current tree age distribution during the decoding process to generate the monthly tree age prediction raster map for the next 12 months in an autoregressive manner. The prediction result is fed back to the boundary prediction model in step S4 in real time, making the boundary division forward-looking. For example, when it is predicted that there will be a continuous drought in the next 3 months, the boundary prediction model will expand the scope of the mechanical operation restriction zone in advance to protect the fragile soil.

[0040] Meanwhile, the feature fusion in this step specifically employs a multimodal cross-attention fusion network. The network's input includes three modalities: temporal meteorological features (dimension T×M, where T is the time series length and M is the number of meteorological variables), and terrain features (dimension H×W×C). topo C topoThe system uses a topographic feature map (number of topographic factors) and a tree age distribution map (dimension H×W×1), where H is the height of the spatial grid and W is the width of the spatial grid. The fusion process is as follows: First, the topographic features and tree age distribution map are projected into a unified embedding space using a 1×1 convolution. Then, global average pooling of the time dimension is applied to the temporal meteorological features to obtain a meteorological embedding vector. Using the meteorological embedding vector as the query and the topographic features as the key and value, cross-attention encoding is performed to generate an attention weight map reflecting "which topographic features are more susceptible to stress under current meteorological conditions." Finally, this attention weight map is pixel-wise weighted and summed with preset stress factor base values ​​(including drought stress coefficient, low-temperature freezing risk coefficient, and topographic drainage index) to generate a comprehensive environmental stress coefficient map (Env). Stress The values ​​range from 0 to 1, with higher values ​​indicating stronger stress. Finally, the environmental stress coefficient map and the tree age spatial distribution map (after normalization) are concatenated along the channel dimension, and then adaptively weighted and fused through a 3×3 convolutional layer to generate the final comprehensive feature tensor F. fused (Dimensions are H×W×C) fused C fused In this embodiment, C represents the number of feature channels after fusion. fused (Take 32), this tensor simultaneously represents the tree age level, topographic features, and meteorological environmental stress at each spatial location.

[0041] In the multi-level dynamic control boundary deduction step, the Non-dominated Sorting Genetic Algorithm II (NSGA-II) is used as the optimization engine for the boundary deduction model, with the comprehensive feature tensor F fused As input, the optimal spatial boundaries of the three types of control areas are derived on a two-dimensional plane of the orchard. The specific implementation method is as follows: (1) Optimization of variable encoding: The orchard plane is discretized into N candidate sub-regions (each sub-region is a 0.5m × 0.5m grid cell). Each individual (chromosome) is encoded as an integer array of length N, with each position taking the value {0, 1, 2}, representing the core growth zone, buffer regulation zone, and mechanical operation restriction zone, respectively. The population size is set to 200, the maximum number of iterations is 500, the crossover probability is 0.8, and the mutation probability is 0.05.

[0042] (2) Definition of biobjective function: Ecological adaptability objective function (minimization direction): Define f eco= w1×(1-light energy capture efficiency) + w2×(1-soil moisture availability) + w3×(1-carbon sink accumulation). Wherein, light energy capture efficiency is calculated using a radiative transfer model (a simplified version of the PROSAIL model) based on the aspect, slope, and tree age (affecting the canopy leaf area index) of each grid cell; soil moisture availability is obtained based on a weighted fusion of the topographic moisture index (TWI) and the meteorological drought index SPEI; carbon sink accumulation is estimated based on tree age and species allometric growth equations. eco The lower the value, the more efficient the light energy utilization, the better the water conditions, and the stronger the carbon sequestration capacity of the orchard under the zoning scheme. w1, w2, and w3 are the weight coefficients of each ecological sub-objective. In this embodiment, the weights for the hilly citrus orchard are set as w1 = 0.3 (light energy), w2 = 0.4 (water), and w3 = 0.3 (carbon sink). Among them, w2 has the highest weight because water is the biggest ecological limiting factor in hilly areas.

[0043] The objective function for operational economic efficiency (direction of minimization): Define f econ = v1 × Turning radius loss during agricultural machinery operation + v2 × Fuel consumption rate + v3 × Harvesting and transportation distance cost. Where, turning radius loss is calculated based on the irregularity of the zone boundaries (boundary fractal dimension) and the area of ​​the region; fuel consumption rate is estimated based on the average slope of the region and the length of the operating path; harvesting and transportation distance cost is calculated by weighting the distances from the center of each region to the main road network of the orchard. econ This is a comprehensive economic cost index for agricultural machinery operations. The lower the value, the more fuel-efficient, time-saving, and labor-saving the agricultural machinery operation is under this zoning scheme. v1, v2, and v3 are the weighting coefficients of each economic sub-objective. In this embodiment, they are set as v1 = 0.4 (turning loss), v2 = 0.3 (fuel consumption), and v3 = 0.3 (transportation cost). Among them, turning loss has the highest weight because the time and fuel wasted on turning at the edge of the field in hilly orchards often account for a large proportion.

[0044] (3) Constraints: The core growth area must meet the following conditions: tree age ≥ 4 years and ≤ 25 years (peak fruiting period), slope ≤ 15°, and environmental stress coefficient ≤ 0.3; the mechanical operation restriction area must meet the following conditions: slope ≥ 25° or environmental stress coefficient ≥ 0.7 or tree age spatial variation coefficient (CV) ≥ 0.4; the remaining areas are classified as buffer adjustment areas.

[0045] (4) Optimization and Decision-Making: After the NSGA-II iteration is completed, a set of Pareto Front non-dominated solutions is obtained. TOPSIS (Topology of Ideal Solutions) is used to select the optimal compromise solution from the Front solution set: the relative proximity of each solution to the ideal optimal solution and the ideal worst solution is calculated, and the solution with the largest proximity is selected as the final output multi-level dynamic control boundary configuration. This boundary configuration is output in the form of three-dimensional coordinate data (category labels for each raster) and a visualized partition boundary vector map (Shapefile format) is generated.

[0046] In the differentiated area management strategy generation and instruction mapping step, based on the multi-level dynamic management boundary vector map output by S4, this step generates specific differentiated management strategies and transforms them into an instruction set executable by the orchard operation terminal: (1) For the core growth area (high-value area during the peak fruiting period): Generate a variable rate fertilization prescription map. Specifically, combining the tree age spatial distribution map (the closer the tree is to the peak fruiting period, the greater the fertilizer requirement) and the basic soil nutrient content (obtained through grid sampling laboratory testing), the recommended application rates of nitrogen, phosphorus, potassium and trace elements for each grid cell are calculated using a Neural Network regression model. The prescription map is output in GeoTIFF format and imported into the vehicle terminal control system of the variable fertilizer applicator (such as the Lovol Arbos variable fertilizer applicator). During operation, the speed of the fertilizer discharge shaft is automatically adjusted through GNSS positioning.

[0047] (2) For the buffer adjustment zone: Generate a variable spraying flight path planning map. Reduce the amount of spraying by 30% to 50% for young fruit trees (tree age < 4 years) and focus on preventing and controlling diseases of the main trunk for aging fruit trees (tree age > 25 years). The flight path planning adopts full coverage path planning based on the A* algorithm to ensure that the operation coverage of spraying drones or self-propelled sprayers reaches more than 95% and the over-spraying rate is less than 5%.

[0048] (3) For areas with restricted mechanical operations: DJI T40 or T50 agricultural drones are used for operations. A terrain-following flight obstacle avoidance path for the drone is generated based on boundary constraints: Input the boundary polygons of the zone and the internal DEM data, and use the Fast Marching Method to generate a strip-shaped sweeping path. Simultaneously, RTK real-time differential positioning is used to achieve terrain-following flight (maintaining a constant altitude of 2-3m above the canopy top), and the trigger distance of the obstacle avoidance radar (such as front and rear phased array radars) is set to 5m. Hover commands are inserted at waypoints along the terrain-following flight path, and variable nozzles are used to achieve precise hovering spraying (the flow rate is reduced to 60% of normal flight speed during hovering to prevent pesticide drift).

[0049] (4) Generation of irrigation valve zone control timing table: Based on the boundary zone division, an on / off timing table is generated for the solenoid valves (installed at the inlet of the drip irrigation branch pipe, such as the DC series solenoid valves from Netafim, Israel) in each zone. A higher soil moisture lower limit threshold is set for the core growth zone (e.g., irrigation is started when the soil water potential is below -30 kPa), a medium threshold is set for the buffer regulation zone (-50 kPa), and the irrigation frequency is further reduced in the mechanical operation restriction zone to reduce the risk of surface runoff and waterlogging. The timing table is sent to the orchard irrigation controller for execution via the Modbus protocol.

[0050] In the closed-loop evaluation and online incremental learning steps of the control effect, after each control strategy is implemented (with one growing season as an evaluation cycle), remote sensing images and sensor feedback data for the next time series (current year or the same period of the following year) of the orchard are acquired. The specific operation is as follows: (1) Calculation of actual growth response indicators: The crown area, normalized vegetation index (NDVI), and crown temperature (extracted from thermal infrared images) of each fruit tree were re-extracted using newly acquired remote sensing images, and the growth of fruit trees in each zone (represented by the annual growth rate of crown area) and yield data (obtained by weighing during harvesting in each zone) were statistically analyzed.

[0051] (2) Deviation calculation: The actual observed values ​​are compared with the corresponding index values ​​predicted by the simulation and inference layer in the digital twin scenario before the control strategy is implemented, and the relative deviation δ = |measured value - predicted value| / predicted value is calculated. When δ exceeds the preset tolerance threshold (set to 15% in this embodiment), it is determined that model correction needs to be triggered.

[0052] (3) Online Incremental Learning and Boundary Recalibration: After triggering, the system organizes the deviation data into paired samples of "current comprehensive feature tensor - actual growth response deviation" according to spatial location. It then uses the Online Stochastic Gradient Descent (SGD) algorithm to fine-tune and update the parameters of the last fully connected layer of the tree age regression prediction model in S2. At the same time, it performs small-batch incremental updates on the attention weights of the multimodal cross-attention fusion network in S3. After the model parameters are updated, the boundary inference process of S4 is re-executed to generate new multi-level dynamic control boundaries. The updated control strategy is automatically output to the operation terminal, realizing continuous optimization and adaptive updates of the entire system during long-term operation. Example 2

[0053] This embodiment focuses on the adaptive adjustment of the visual detection model under different orchard scenarios. For non-citrus fruit trees (such as apples, pears, peaches, etc., which have different tree structures), the visual detection model in step S2 is adaptively modified according to the specific tree species characteristics. For example, for apple trees with a central trunk or Y-shape, a main trunk direction detection submodule can be added to the edge refinement branch to distinguish between the main trunk and main branches; for vine fruit trees such as grapes, the "single tree" detection can be adjusted to "single row of trees" detection, and the tree age regression model can be adjusted to row unit tree age mean prediction accordingly. Example 3

[0054] This embodiment focuses on the boundary deduction implementation method of reinforcement learning instead of genetic algorithms. In another preferred embodiment, the boundary deduction model in step S4 can use the Deep Deterministic Policy Gradient (DDPG) algorithm instead of NSGA-II, which is suitable for scenarios where the orchard environment changes frequently and requires real-time dynamic adjustment of the boundary. Specifically, the comprehensive feature tensor is regarded as the environmental state, the spatial distribution of boundary categories is regarded as a continuous action space (the discrete category selection is relaxed to a continuous probability distribution through Gumbel-Softmax), and the reward function is set as the improvement of the ecological adaptation goal minus the increment of the economic cost of operation. The agent interacts with the digital twin scene, and through iterative training of the actor-critic network, when the reward value converges, the trained policy network can be directly used to make a rapid response to new meteorological input data (such as typhoon or cold wave weather forecasts), and output the updated control boundary within seconds, which is more suitable for emergency scenarios. This implementation method has the same inventive concept as Embodiment 1 and both fall within the protection scope claimed by this invention. Example 4

[0055] This embodiment focuses on the adjustability and adaptability of threshold parameters. The various preset thresholds involved in this invention (such as slope thresholds of 15° and 25°, environmental stress coefficient thresholds of 0.3 and 0.7, tree age segments of 4 years and 25 years, deviation tolerance threshold of 15%, and canopy width change rate calculation interval of 15 days, etc.) are all optimized parameters based on typical hilly citrus orchards. In practical applications, these thresholds are optimized and adjusted through orthogonal experiments or response surface methodology according to the specific orchard's terrain undulation, fruit tree variety growth characteristics, local climate characteristics, and agricultural machinery equipment parameters to achieve the best boundary delineation effect and economical management. All such threshold adjustments do not depart from the basic technical framework of this invention and should be considered as direct equivalent implementations of this invention. Example 5

[0056] This embodiment focuses on system integration and hardware deployment. As a complete hardware system integration solution of this invention, it includes: a data acquisition layer (drone, weather station, sensor network), a data processing layer (high-performance server or cloud platform configured with NVIDIA GPU, deploying the deep learning framework PyTorch and the digital twin engine Unity 3D), a network transmission layer (5G / 4G and Wi-Fi 6 hybrid networking), and an execution layer (variable fertilizer applicator, variable sprayer, drone, intelligent irrigation controller). The layers communicate loosely via the MQTT IoT protocol and RESTful API interfaces, facilitating system maintenance and functional expansion. Example 6

[0057] This embodiment focuses on a tree age regression prediction model based on graph neural networks.

[0058] For constructing the feature vector of a single fruit tree, the system builds an initial feature description vector for each fruit tree identified by the visual detection model in the orchard. This vector consists of eight dimensions of information: First, the crown diameter extracted by a deep learning instance segmentation network, reflecting the horizontal extension of the crown; second, the crown projection area, i.e., the area covered by the crown under vertical projection; third, the trunk diameter, measured at a fixed height above the ground, which is an important structural parameter reflecting the growth of the fruit tree; fourth, the branch density, obtained by statistically analyzing the edge response intensity within the segmentation mask area, characterizing the density of the internal structure of the crown; fifth, the crown change rate, calculated by comparing the change in crown area of ​​the same tree in two adjacent remote sensing observations and dividing by the time interval, reflecting the recent actual growth vitality of the fruit tree; sixth to eighth, the terrain slope, aspect, and altitude values ​​of the location of the fruit tree, respectively. These eight features together constitute an eight-dimensional initial feature vector, serving as the basic input for subsequent tree age prediction.

[0059] For the spatial message passing mechanism of graph neural networks, all the fruit trees in the entire orchard are regarded as a spatial graph network structure, where each fruit tree is a node in the graph. The connection relationship between nodes is determined by spatial distance: when the actual distance between two fruit trees is less than a preset neighborhood radius, an undirected edge is established between them, and the weight of the edge is the reciprocal of the distance between the two trees. The closer the fruit trees are, the greater their mutual influence in information transmission. In each layer of the graph neural network, each node aggregates the feature information of all its neighboring nodes. The aggregation method is as follows: after linearly transforming the feature vector of each neighboring node, multiplying it by its edge weight and dividing it by the geometric mean of the degree of the two endpoints to achieve normalization, and then summing the contributions of all neighbors to obtain the aggregated message vector of the node. Subsequently, this aggregated message vector is fused with the node's own terrain covariate features (i.e., slope, aspect, and altitude, three fixed environmental factors) through a trainable weight matrix and processed by a nonlinear activation function to obtain the updated feature vector of the node in the next layer. This process iterates multiple times, ensuring that the final feature representation of each fruit tree not only includes its own morphological information, but also fully absorbs the spatial coordination information of neighboring fruit trees and the systematic influence of the terrain environment.

[0060] In the core design of using topographic covariates as spatial bias terms, topographic factors (slope, aspect, and altitude) are always injected into the network as independent bias terms in each layer update of the graph neural network, rather than simply as part of the initial input. This means that even in the high-level abstract features of the network, topographic information still directly participates in feature evolution, forcing the model to maintain sensitivity to topographic differences throughout the learning process. For example, the model can automatically learn that, for the same canopy size, fruit trees on sunny slopes are often younger than those on shady slopes with the same canopy size due to sufficient sunlight but high water evaporation; and that fruit trees in high-altitude areas grow slowly due to lower temperatures, and therefore are usually older for the same canopy size than those in low-altitude areas. This design, which embeds prior topographic knowledge into the network architecture as a hard bias, effectively eliminates the interference of topographic heterogeneity on the empirical relationship between canopy size and tree age, significantly improving the generalization ability of tree age estimation under complex topographic conditions.

[0061] For the spatial smoothing regularization in the loss function, the loss function consists of two terms when training this graph neural network. The first term is the conventional prediction error term, which is the average of the squares of the differences between the predicted and actual tree ages of all training samples. The second term is the spatial smoothing regularization term, which is calculated as follows: for each connected pair of adjacent fruit trees in the graph, calculate the square of the difference between their predicted ages, multiply it by the weight coefficient of that edge (the reciprocal of the distance), and finally sum over all edges. The physical meaning of this regularization term is that spatially adjacent fruit trees should not have drastic changes in age due to similar growing environments. This constraint ensures that the final output spatial distribution map of tree age remains smooth and continuous overall, while also preserving the differences in actual tree age caused by replanting, thinning, etc. in local areas through the prediction error term.

[0062] For generating a full-coverage spatial distribution map of tree age, since visual detection models cannot cover every grid location in the orchard (such as gaps between rows, missing trees, etc.), this invention employs a Kriging interpolation method with external drift to generate a full-coverage raster map of tree age distribution for undetected spatial areas. The core idea of ​​this method is that, in addition to using the known tree age values ​​of detection points for spatial interpolation, topographic aspect and elevation are introduced as auxiliary external drift variables into the interpolation model. Specifically, during the interpolation process, the estimated value of unknown points is obtained by a weighted combination of two parts: one part is the linear weighted sum of the known tree age values ​​of detection points, and the other part is the linear combination of the external drift function (i.e., topographic factors) at the unknown points. In this way, the interpolation result not only maintains the continuous spatial transition of tree age but also fully reflects the systematic influence of topography on tree age distribution, thereby generating a high-precision spatial distribution map of tree age that covers the entire orchard and has a spatial resolution consistent with remote sensing imagery. Example 7

[0063] This embodiment focuses on multimodal cross-attention fusion networks and the calculation of environmental stress coefficients.

[0064] For feature projection and a unified embedding space, the terrain feature map (containing four channels: slope, aspect, elevation, and surface roughness) and the tree age distribution map (single channel) have the same resolution and number of rows and columns in the spatial dimension. First, these two types of data are projected into a unified feature embedding space through pointwise convolution (i.e., 1×1 convolution). The purpose of this projection process is to map data of different modalities and physical dimensions to vector representations of the same dimension, enabling the subsequent attention mechanism to compute the correlation between locations at comparable scales.

[0065] For cross-attention encoding using meteorological features as query vectors, time-series meteorological data (including historical daily temperature, precipitation, solar radiation, wind speed, etc.) is compressed into a fixed-length meteorological embedding vector after global average pooling along the time dimension. This vector serves as the query vector for the subsequent cross-attention mechanism. In cross-attention calculation, the vector projected from each spatial location in the topographic feature map is used as both the key vector and the value vector. The attention score is calculated by performing a dot product operation between the meteorological query vector and the topographic key vector at each location, then scaling by the square root of the embedding dimension, and finally normalizing it to a probability distribution form of attention weights using a softmax function. The physical meaning of this weight is the sensitivity of each topographic location in the orchard to meteorological stress under current meteorological conditions (such as high temperature and drought, heavy rainfall and low temperature). A higher attention weight indicates that the location is more susceptible to environmental stress under current meteorological conditions. Subsequently, these attention weights are used as coefficients to perform a weighted summation of the topographic value vectors, resulting in the feature representation enhanced by the meteorological-topographic interaction.

[0066] To generate the comprehensive environmental stress coefficient, the system performs a weighted synthesis of three basic stress factors under cross-attention weight modulation. The first factor is the drought stress coefficient, derived from the standardized precipitation evapotranspiration index. This index comprehensively reflects the long-term impact of precipitation and temperature on soil moisture balance; a lower index value indicates a more severe drought. After mapping transformation, it is converted into a drought stress intensity value between 0 and 1. The second factor is the low-temperature freezing damage risk coefficient, obtained by normalizing the frequency of extreme low-temperature events in the historical records of the location. The third factor is the topographic catchment risk coefficient, quantified using the topographic humidity index. This index is calculated based on catchment area and slope; a higher value indicates a greater likelihood of water accumulation at the location. The three basic stress factors are not simply averaged but their respective weights are dynamically adjusted according to current meteorological conditions: in drought years, the meteorological-topographic attention weight tends to increase the contribution of the drought stress factor; in years with continuous rain, the contribution of the catchment risk factor is increased. Finally, the three weighted stress factors are mapped to the interval of 0 to 1 through a nonlinear compression function to obtain a comprehensive environmental stress coefficient map, where 0 represents the optimal environment with no stress and 1 represents the most severe stress.

[0067] For the final generation of the comprehensive feature tensor, the normalized tree age distribution map and the comprehensive environmental stress coefficient map are concatenated along the channel dimension to form a dual-channel intermediate feature map. This intermediate feature map is then processed by a 3×3 spatial convolutional layer, combined with batch normalization and a nonlinear activation function, to output the final comprehensive feature tensor. This tensor has 32 channels, meaning each spatial location uses a 32-dimensional feature vector to comprehensively represent the tree age level, terrain attributes, and meteorological stress coupling information at that point. These 32 abstract channels do not have artificially assigned physical meanings but are learned autonomously by the network and are used as the optimal feature representation for the subsequent boundary extrapolation model input. Each channel may represent a specific combination of growth potential or limiting factors. Example 8

[0068] This embodiment focuses on multi-objective boundary deduction and optimization algorithms.

[0069] To define the optimization decision variables, the entire orchard is divided into several square grid cells according to spatial resolution. The category of each grid cell is the core optimization decision variable. Each variable is limited to one of three discrete values, representing whether the grid cell is assigned to the core growth zone, the buffer zone, or the mechanical operation restriction zone. Therefore, a complete orchard zoning scheme is a combination of all grid category values, constituting a candidate solution to a large-scale discrete optimization problem.

[0070] The ecological adaptability objective function is calculated to assess the degree to which a zoning scheme protects the growth potential of fruit trees; a smaller objective value indicates better ecological adaptability. This objective is composed of a weighted sum of three sub-terms, each expressed as "1 minus the positive indicator," to unify the optimization direction towards minimization.

[0071] The first sub-term is the loss term for light energy capture efficiency. The calculation of light energy capture efficiency is based on the following: the leaf area index (LAI) at each grid location is estimated from the tree age at that location using the allometric growth equation. The spatial distribution of the LAI is influenced by slope aspect—leaf aggregation is typically lower on sunny slopes and higher on shady slopes. Substituting the LAI into the simplified radiative transfer model allows us to calculate the fraction of photosynthetically active radiation absorbed at that location. The higher this fraction, the more efficient the light energy utilization, and the smaller the corresponding loss term (1 minus the fraction).

[0072] The second sub-item is the loss of soil moisture availability. This indicator is the average of the normalized topographic moisture index and the standardized precipitation evapotranspiration index. The topographic moisture index reflects the soil moisture replenishment advantage brought by the catchment topography and is a static topographic attribute; the standardized precipitation evapotranspiration index reflects the recent meteorological dryness and wetness conditions and is a dynamic meteorological attribute. The average of the two reflects the water supply capacity of this location under current and long-term conditions.

[0073] The third sub-item is the loss of carbon sequestration. Based on tree age and trunk diameter, aboveground biomass is estimated using the allometric growth equation for this tree species, then multiplied by the carbon conversion coefficient to obtain the carbon storage per unit area. After normalization, this becomes the carbon sequestration index. The higher this value, the stronger the carbon sequestration capacity of the area and the greater its ecological value. The sum of these three sub-items, each assigned a weighting coefficient, yields the comprehensive ecological adaptation cost of this zoning scheme.

[0074] The objective function for calculating the economic efficiency of a work operation aims to evaluate the impact of a zoning scheme on the efficiency and economic cost of mechanized operations. A smaller objective value indicates a more economical and efficient operation. This objective is also composed of a weighted sum of three sub-items.

[0075] The first sub-item is the turning radius loss during agricultural machinery operation. This loss is quantified by calculating the tortuosity of the boundary of each zone: the larger the ratio of the perimeter to the area of ​​the zone boundary, the narrower or more irregular the shape of the area, and the more frequently the agricultural machinery needs to turn and maneuver during operation, thus increasing the time spent ineffectively and fuel consumption. This ratio is multiplied by the minimum turning radius parameter of the agricultural machinery to obtain the turning loss value for that zone.

[0076] The second sub-item is fuel consumption rate. Fuel consumption for each grid position is obtained by multiplying the base fuel consumption by the slope correction factor and then by the work path length. The slope correction factor is exponential; fuel consumption increases proportionally with each increase in slope angle, reflecting the additional fuel consumption caused by increased engine load during uphill operations. The work path length is estimated from the area of ​​the region and the preset work width.

[0077] The third sub-item is the cost of harvesting and transporting distances. This cost equals the shortest distance from each zone center to the orchard's main road network multiplied by the region's estimated yield. The estimated yield is predicted by the tree age using a yield model. This design assigns higher transport costs to areas farther from the main road but with higher yields, thus favoring the reclassification of these areas into less intensive management zones or the addition of temporary transport routes during optimization.

[0078] For the handling of hard constraints and penalty functions, the optimization process must satisfy a series of hard constraints. Constraint 1: Any grid cell included in the core growth zone must simultaneously meet three conditions: tree age within the peak fruiting period, terrain slope less than a preset upper limit, and environmental stress coefficient less than a lower limit. All three are indispensable. Constraint 2: Any grid cell included in the mechanized operation restriction zone will be forcibly included if it meets any one of the following conditions: slope exceeding a preset upper limit, environmental stress coefficient exceeding an upper limit, or local tree age spatial variation coefficient exceeding a threshold. The tree age spatial variation coefficient refers to the ratio of the standard deviation to the mean of the tree age values ​​of each grid cell within a local window centered on that grid cell. An excessively high coefficient indicates that the fruit tree ages in the area are mixed and the growth is uneven, making it unsuitable for large-scale mechanized unified operation. For solutions that do not meet the above constraints, the optimization algorithm uses a penalty function method: a very large penalty term is superimposed on the objective function value, causing solutions that violate the constraints to be naturally eliminated during the evolutionary process.

[0079] For the solution process of the Non-Dominated Sorting Genetic Algorithm (NSGA-II), the NSGA-II algorithm is used to solve the above bi-objective optimization problem. First, a certain number of initial partitioning schemes are randomly generated as the first generation population. During each generation of evolution, the ecological fitness objective value and the operational economy objective value are calculated for each scheme in the population. Then, non-dominated sorting is performed based on these two objective values: if a scheme is not inferior to another scheme in either objective and is superior to the other in at least one objective, then the former dominates the latter. All schemes not dominated by any other scheme form the first frontier, which is removed from the population. The remaining schemes are then sorted to obtain the second frontier, and so on. Simultaneously, the crowding distance of each scheme is calculated to measure its sparsity on its frontier. In the selection phase, schemes with smaller frontier numbers and larger crowding distances are preferentially selected to enter the next generation to maintain both population diversity and convergence. After a predetermined number of iterations, the population converges to a set of Pareto optimal solutions, which form different compromise schemes between the two objectives.

[0080] For the optimal compromise solution selection (TOPSIS method), when choosing the final implementation scheme from the Pareto front solution set, the approximation-to-ideal-solution ranking method is adopted. This method first finds the combination of the minimum ecological objective and the minimum economic objective among all solutions, forming a virtual ideal optimal solution; simultaneously, it finds the combination where both objectives are maximized, forming a virtual ideal worst solution. Then, it calculates the distance from each actual solution in the front to the ideal optimal and ideal worst solutions, and ranks each solution according to its relative proximity. The closer the proximity is to 1, the closer the solution is to the ideal optimal solution and the further away it is from the ideal worst solution, indicating the most balanced compromise scheme. This scheme is ultimately output as a multi-level dynamic management boundary configuration for the orchard.

[0081] For the implementation logic of reinforcement learning alternatives, in application scenarios requiring real-time response to sudden weather changes, this invention provides an alternative implementation scheme using a deep deterministic policy gradient algorithm. In this scheme, the comprehensive feature tensor is encoded by a deep convolutional neural network to form an environmental state vector. Optimized actions are no longer discrete category encodings, but rather the discrete category selection is relaxed to a continuous probability distribution using Gumbel-Softmax technology, making the action space in reinforcement learning continuous. The reward function is directly defined as the improvement in the ecological adaptability objective minus the increment of the operational economic objective. The agent trains the actor network to output the optimal action through continuous interaction with the digital twin scene, and the critic network evaluates the value of the action. After training convergence, new meteorological data input can drive the system to output updated boundary delineation within seconds, making it particularly suitable for rapid response in sudden weather warning scenarios such as typhoons and cold waves. Example 9

[0082] This embodiment focuses on algorithmic extensions in the generation of control strategies.

[0083] For the algorithm generating the variable-rate fertilization prescription map, the recommended nitrogen application rate per unit area in the core growth zone is calculated by a mechanism-data dual-driven model. This model uses the recommended base fertilization rate for the variety as a benchmark, multiplied by two adjustment coefficients. The first adjustment coefficient is the tree age adjustment term: when the tree age at this location is close to the peak fruiting period, the adjustment coefficient is close to 1, and the fertilization rate remains at the benchmark level; when the tree age is below the peak fruiting period, the adjustment coefficient is less than 1, and fertilization is reduced to avoid excessive vegetative growth in young trees; when the tree age is above the peak fruiting period, the adjustment coefficient gradually decreases, reflecting the reduced nutrient absorption and utilization efficiency of aging fruit trees. The second adjustment coefficient is the soil organic matter adjustment term: when the soil organic matter content at this location is higher than the optimal level, the adjustment coefficient is less than 1, and fertilization is appropriately reduced to utilize the soil's own nutrient mineralization capacity; when the organic matter content is lower than the optimal level, the adjustment coefficient is greater than 1, and fertilization is increased to compensate for insufficient basic soil fertility. Multiplying the two adjustment coefficients by the base fertilization amount yields the recommended fertilization amount for that grid position, ultimately generating a fully covered variable fertilization prescription map.

[0084] For the UAV terrain-following flight path planning algorithm, the path planning problem in the restricted mechanical operation area is modeled as a shortest time path problem with terrain and obstacle constraints. The system first constructs a velocity field function based on terrain slope and abrupt changes in canopy height density: the steeper the slope and the denser the obstacles, the lower the allowed flight speed for the UAV. Then, a fast-travel method is used to solve for the arrival time field from the starting point to the destination. The core physical model of this algorithm is similar to wave propagation in a non-uniform medium—the wavefront advances rapidly in high-velocity areas and expands slowly in low-velocity areas. After obtaining the arrival time field for the entire area, the optimal flight path from the starting point to the destination is obtained by tracing backward along the direction of the fastest descent of the time field gradient from the destination position. This path automatically avoids areas with high slope and high obstacle density during planning, while strictly confining itself within the boundary of the restricted mechanical operation area, preventing it from crossing into other controlled areas. At critical waypoints on the path, the system automatically inserts hovering commands, enabling the UAV to hover stably above the target position for precise spraying. During hovering, the spraying flow rate is appropriately reduced to prevent pesticide drift. Example 10

[0085] This embodiment focuses on the closed-loop online incremental learning algorithm.

[0086] For the quantitative assessment of deviation indicators, in the evaluation cycle after the completion of each control strategy, the system reacquires the latest remote sensing images and IoT sensor data of the orchard, and re-extracts actual growth response indicators such as the annual canopy growth rate, normalized difference vegetation index, and actual yield of each zone for each fruit tree. These measured values ​​are then compared item by item with the predicted values ​​given in the digital twin scenario before the implementation of the control strategy, and the relative deviation of each item is calculated, that is, the absolute value of the difference between the measured value and the predicted value divided by the predicted value. The larger the relative deviation, the more serious the deviation between the actual growth response and the simulation prediction.

[0087] Regarding triggering conditions and online gradient updates, the system automatically triggers the online incremental learning mechanism when the relative deviation of any of the above indicators exceeds a preset tolerance threshold. For the tree-age regression graph neural network, mini-batch online stochastic gradient descent updates are mainly performed on the weight parameters of the last fully connected layer. The online learning rate decays proportionally with the number of update rounds. This decay strategy is designed to ensure that the model absorbs the corrective information brought by new data while preventing drastic overwriting of old knowledge and avoiding catastrophic forgetting problems.

[0088] To protect existing knowledge, this invention introduces an elastic weight consolidation constraint mechanism for projected weights in multimodal cross-attention fusion networks. The core idea of ​​this mechanism is to apply different levels of protection to each network parameter during incremental learning. The degree of protection is based on the diagonal elements of the Fisher information matrix, which reflects the importance of each parameter's contribution to the final prediction accuracy during training on the old task—parameters with higher importance are allowed smaller changes during updates. Specifically, an additional penalty term is added to the loss function of incremental learning: this term equals the square of the difference between the current value and the optimal value on the old task for each parameter, multiplied by the corresponding Fisher information value, and then multiplied by a global consolidation strength coefficient. This design allows the model to selectively adjust the parameters that contribute the most to the current bias when correcting using new biased data, while effectively protecting parameters that play a crucial role in maintaining the overall model's stability and generalization ability, thus achieving an optimal balance between adapting to new data and preserving existing knowledge.

[0089] For boundary recalibration and system self-evolution, after incrementally updating the model parameters, the system automatically uses the updated tree age prediction model and fusion network to regenerate the tree age distribution map and comprehensive feature tensor, and re-executes the boundary inference optimization process, outputting a new round of multi-level dynamic control boundaries. The new control strategy is then distributed to each operational terminal for execution. Through this complete closed loop of "execution-evaluation-deviation triggering-incremental learning-recalibration-re-execution," the entire system can continuously track the actual growth changes and environmental fluctuations of the orchard during long-term operation, achieving continuous optimization and self-evolution of the control strategy, and always maintaining an accurate mapping of the orchard's true state and optimal decision-making capability.

[0090] The above specific embodiments are merely several preferred embodiments of the present invention. Based on the technical solutions of the present invention and the relevant teachings of the above embodiments, those skilled in the art can make various alternative improvements and combinations to the above specific embodiments.

Claims

1. A method for orchard area boundary identification and management based on meteorological topography and tree age characteristics, characterized by the following steps: include: S1: Acquire multi-source heterogeneous data of the target orchard, including time-series meteorological data, topographic digital elevation model data, and orchard remote sensing image data containing individual morphological characteristics of fruit trees; S2: Based on deep learning network, a visual detection model is built to extract the crown and trunk features of each fruit tree from orchard remote sensing image data. Combined with time-series meteorological data and topographic digital elevation model data, the individual tree age value and tree age spatial distribution map of each fruit tree are output through the tree age regression prediction model. S3: Construct a digital twin scene that couples meteorology, topography, and tree age, and perform spatiotemporal registration and feature fusion of time-series meteorological data, topographic digital elevation model data, and tree age spatial distribution map to generate a comprehensive feature tensor that reflects the differences in growth potential and environmental stress in different areas of the orchard. S4: Input the comprehensive feature tensor into the boundary inference model constructed based on reinforcement learning or genetic optimization algorithm, with ecological adaptability and operational economy as dual objective constraints, dynamically infer and output the multi-level dynamic control boundary inside the orchard. The multi-level dynamic control boundary includes the core growth zone, buffer regulation zone and mechanical operation restriction zone. S5: Based on multi-level dynamic control boundaries and combined with real-time or forecast meteorological data, generate differentiated regional control strategies and map the control strategies to the execution instruction set of the orchard operation terminal to realize orchard zoned control; The feature fusion in step S3 employs a multimodal cross-attention fusion network. This network performs cross-attention encoding on temporal meteorological features and topographic features to generate an attention weight map. Then, the attention weight map is summed pixel-by-pixel with preset stress factor base values ​​to generate a comprehensive environmental stress coefficient map. Finally, the environmental stress coefficient map is concatenated with the tree age spatial distribution map along the channel dimension, and then subjected to an adaptive weighted fusion through a 3×3 convolutional layer to generate the final comprehensive feature tensor. The preset stress factor base values ​​include drought stress coefficient, low temperature freezing risk coefficient, and topographic drainage index.

2. The orchard area boundary identification and management method based on meteorological topography and tree age characteristics according to claim 1, characterized in that: The time-series meteorological data in step S1 includes historical daily temperature, precipitation, solar radiation, wind speed, and weather forecast data within the future preset time window; the topographic digital elevation model data includes slope, aspect, altitude, and surface roughness factor; and the orchard remote sensing image data includes multispectral imagery, thermal infrared imagery, and visible high-resolution imagery.

3. The orchard area boundary identification and management method based on meteorological topography and tree age characteristics according to claim 1, characterized in that: The visual detection model in step S2 uses the Mask R-CNN instance segmentation network and the backbone network uses SwinTransformer Tiny. On the basis of the backbone network, an attention enhancement module and an edge refinement branch are embedded to extract the crown diameter, crown projection area, trunk diameter and branch distribution density of a single fruit tree. At the same time, the crown change rate is calculated by using the change in crown area of ​​the same single tree detected in two adjacent time series images.

4. The orchard area boundary identification and management method based on meteorological topography and tree age characteristics according to claim 1, characterized in that, The tree age regression prediction model is constructed as follows: the crown width and trunk features extracted by the visual detection model are used as inputs, combined with the historical planting year or standard age sample labels of the corresponding fruit trees, and a graph neural network is used for training; during the training process, the slope aspect and altitude factors in the topographic digital elevation model data are introduced as spatial covariates to eliminate the influence of topographic heterogeneity on the tree age-crown width growth relationship, and the individual tree age value is corrected based on the crown width change rate of adjacent time series images.

5. The orchard area boundary identification and management method based on meteorological topography and tree age characteristics according to claim 1, characterized in that: The digital twin scenario includes a physical entity layer, a data mapping layer, and a simulation and inference layer. The physical entity layer is used to access orchard IoT sensor data in real time. The data mapping layer is used to construct a three-dimensional virtual orchard model and overlay a spatial distribution map of tree age. The simulation and inference layer is based on a long short-term memory network or a temporal Transformer model to make rolling predictions on the tree age growth trend and boundary drift under different weather scenarios in the future, and feeds the prediction results back to the boundary inference model to iteratively update the boundary.

6. The method for orchard area boundary identification and management based on meteorological topography and tree age characteristics according to claim 1, characterized in that: The boundary extrapolation model in step S4 uses a non-dominated sorting genetic algorithm or a deep deterministic strategy gradient algorithm for multi-objective optimization. The quantitative indicators of ecological adaptability include regional light energy interception efficiency, soil moisture availability and carbon sink accumulation. The quantitative indicators of operational economy include agricultural machinery turning radius loss, fuel consumption rate and harvesting and transportation distance cost. The multi-level dynamic control boundary is determined based on the optimal compromise solution of the Pareto front solution set.

7. The method for orchard area boundary identification and management based on meteorological topography and tree age characteristics according to claim 1, characterized in that: The core growth zone is an area where trees are in their peak fruiting period, the terrain slope is less than the preset threshold, and the environmental stress coefficient is below the lower limit. This area is designated as a priority zone for water and fertilizer input and flower and fruit management. The buffer zone is an area where trees are young or senescent, or where the environmental stress coefficient is within the preset middle range. This area is designated as a zone for dynamically adjusting agronomic measures. The mechanical operation restriction zone is an area where the terrain slope is greater than the preset upper limit, or where there is a risk of water accumulation, or where the tree age distribution is uneven. This area is designated as a zone where the volume of machinery entering is restricted and drone operations are adopted.

8. The method for orchard area boundary identification and management based on meteorological topography and tree age characteristics according to claim 1, characterized in that, The differentiated regional control strategy in step S5 includes: generating variable rate fertilization prescription maps, variable spraying flight path planning maps, and irrigation valve zoning control timing tables based on multi-level dynamic control boundaries; and generating UAV terrain-following flight obstacle avoidance paths and hovering spraying strategies based on boundary constraints for mechanical operation restricted areas.

9. The method for orchard area boundary identification and management based on meteorological topography and tree age characteristics according to claim 1, characterized in that, It also includes step S6: constructing a closed-loop evaluation module for control effect. After each control strategy is executed, it acquires remote sensing images and sensor feedback data of the orchard in the next time series, calculates the deviation between the actual growth response index and the simulation prediction value in the digital twin scenario; when the deviation exceeds the preset tolerance threshold, it triggers the online incremental learning mechanism of the boundary extrapolation model, uses the deviation data to correct the model parameters, and recalibrates and automatically updates the multi-level dynamic control boundary.

Images