An agricultural planting intelligent management method and system based on an internet of things

Abstract

This application provides an intelligent management method and system for agricultural planting based on the Internet of Things (IoT), relating to the field of agricultural planting. It solves the technical problem that existing technologies fail to fully integrate crop physiological state and individual spatial distribution information, making precise and coordinated management of water, fertilizer, and pesticides at the individual plant scale impossible, resulting in resource waste and low control efficiency. The method includes: acquiring environmental data and calculating current irrigation demand, controlling the irrigation execution unit to perform variable irrigation; performing preliminary screening and identification of local pests and diseases in field images, identifying suspected lesion areas, and obtaining images of these suspected lesion areas; generating an infection probability heatmap based on the suspected lesion area images; and planning the optimal flight path of a pesticide application drone based on the infection probability heatmap and a digital twin map of the farmland, and controlling the drone to perform targeted pesticide application. This application is used in the process of agricultural planting management.

Description

Technical Field

[0001] This application relates to the field of agricultural planting, and in particular to an intelligent management method and system for agricultural planting based on the Internet of Things. Background Technology

[0002] With the acceleration of agricultural modernization, precision agriculture technology is receiving increasing attention for improving resource utilization efficiency, ensuring food security, and promoting sustainable development. Traditional agricultural production mainly relies on manual experience for water and fertilizer management and pest and disease control, which suffers from problems such as delayed response, strong subjectivity, and serious resource waste. In recent years, technologies such as the Internet of Things (IoT), remote sensing, and artificial intelligence have been gradually applied to farmland management. Some systems acquire environmental parameters by deploying soil sensors or weather stations and estimate crop water requirements based on the Penman-Monteith model recommended by the FAO, achieving a certain degree of automated irrigation. At the same time, some studies have used drones equipped with multispectral cameras or RGB cameras to collect field images and combined them with convolutional neural networks (CNNs) for disease identification to assist plant protection decision-making.

[0003] However, existing technologies still have obvious shortcomings: current intelligent irrigation and pest control systems generally adopt a "perception-decision" separation architecture, that is, the irrigation module only calculates water demand based on meteorological and soil data, while the plant protection module only judges diseases based on visual images. The two lack a collaborative mechanism and neither fully integrates crop physiological state and individual spatial distribution information.

[0004] Therefore, this application provides an intelligent management method and system for agricultural planting based on the Internet of Things. Summary of the Invention

[0005] This application provides an intelligent management method and system for agricultural planting based on the Internet of Things, which solves the technical problem that the existing technology does not fully integrate crop physiological state and individual spatial distribution information, and cannot achieve precise and coordinated management of water, fertilizer and pesticides at the single plant scale, resulting in resource waste and low control efficiency.

[0006] To achieve the above objectives, this application adopts the following technical solution: Firstly, an intelligent management method for agricultural planting based on the Internet of Things is provided, including: Environmental data is collected in real time through an Internet of Things (IoT) sensor network deployed in farmland. The environmental data includes soil moisture, air temperature and humidity, light intensity, wind speed, rainfall, and crop canopy temperature. Based on the environmental data, and combining the modified Penman-Monteith evapotranspiration model and the Crop Water Stress Index (CWSI), the current irrigation demand is dynamically calculated; and the irrigation execution unit is controlled to perform variable irrigation according to the irrigation demand. The edge computing gateway performs local pest and disease screening on field images to identify suspected lesion areas and obtain images of suspected lesion areas; an infection probability heat map is generated based on the images of suspected lesion areas; the images of suspected lesion areas contain one or more marked suspected lesion areas; Based on the infection probability heatmap and the farmland digital twin map, the optimal flight path of the spraying drone is planned, and the spraying drone is controlled to perform targeted spraying.

[0007] Based on the above technical solutions, this application provides an IoT-based intelligent agricultural planting management method. Addressing the problems of isolated irrigation and plant protection decisions and a lack of comprehensive consideration of crop physiological states and individual spatial distribution in existing technologies, leading to water and pesticide waste and poor control effects, this solution achieves precise operation at the single-plant level through multimodal perception fusion and closed-loop intelligent control. The system utilizes environmental parameters such as soil moisture and canopy temperature, combined with a modified Penman-Monteith model and the Crop Water Stress Index (CWSI), to dynamically assess the actual water requirements of crops, avoiding water supply based on theoretical evapotranspiration under stress conditions. Simultaneously, it uses an edge computing gateway to perform initial disease screening in the field, uploading only images of marked suspected lesion areas, effectively reducing communication overhead and ensuring data security. Based on this, a high-resolution infection probability heatmap is generated. This heatmap is further spatially aligned with a farmland digital twin map integrating geographic coordinates, obstacle distribution, and crop row direction information to plan the optimal flight path covering high-risk areas while meeting obstacle avoidance constraints, guiding the spraying drone to perform targeted spraying. As a result, water, fertilizer and pesticide management has shifted from "uniform field management" to "tailored management based on individual plants," significantly improving resource utilization efficiency, strengthening early warning and precise intervention capabilities, and effectively overcoming key bottlenecks in traditional agriculture such as excessive irrigation, blind pesticide application, and extensive operations, providing efficient, green and scalable technical support for large-scale smart planting.

[0008] In conjunction with the first aspect above, in one possible implementation, the method for obtaining the current irrigation demand includes: Based on the aforementioned soil moisture, air temperature, wind speed, and light intensity, the reference crop evapotranspiration was calculated using the FAO Penman-Monteith formula. ; Based on the crop canopy temperature Using a pre-existing meteorological-canopy temperature boundary model, the theoretical minimum canopy temperature under no water stress was determined. Theoretical maximum canopy temperature under severe stress And calculate the crop water stress index. : ; Obtain the corresponding crop coefficient based on the current crop growth stage. In conjunction with the effective irrigation area per plant Compared with the preset stress sensitivity coefficient The current irrigation demand is calculated using the following formula. : .

[0009] In conjunction with the first aspect above, in one possible implementation, the step of obtaining the corresponding crop coefficient based on the current crop growth stage... ,include: The normalized vegetation index (NDVI) is calculated by collecting real-time visible and near-infrared reflectance data of crop canopy using multispectral imaging sensors deployed in farmland. The cumulative effective accumulated temperature (GDD) output by the soil-crop coupling model is acquired synchronously. The GDD is obtained by dynamic integration of the daily average air temperature and the crop baseline development temperature. The NDVI and GDD are fused to construct a two-dimensional growth phase feature vector. ; The feature vector The data is input into a pre-trained lightweight crop growth stage classifier, which outputs the label of the current sub-growth stage. The classifier is based on the MobileNetV3 architecture with enhanced attention mechanism. The sub-growth stages include, but are not limited to: seedling stage, tillering / branching stage, jointing / budding stage, flowering stage, grain filling / enlargement stage, and maturity stage. Based on the detailed growth stage labels, the corresponding benchmark crop coefficients are dynamically retrieved from the cloud-based crop knowledge graph. The final crop coefficient is obtained by adaptively correcting it based on current environmental stress factors. : ; Among them, environmental stress factors include the current air-water vapor pressure difference. and measured root zone soil volumetric water content ; The optimal VPD threshold for this growth stage; Field holding capacity; Points of wilting in the field; This represents the environmental sensitivity weighting coefficient.

[0010] In conjunction with the first aspect above, in one possible implementation, the effective irrigation area per plant... Methods for obtaining [the information] include: By using multi-view RGB-D cameras deployed in the field or lidar mounted on drones, a three-dimensional point cloud scan of the target crop area is performed to generate a high-precision three-dimensional model of the canopy. Based on the aforementioned 3D model, the improved DBSCAN clustering algorithm was used to separate the spatial coordinates of individual crop plants. ; Based on the crop type and current growth stage, the horizontal root expansion radius is retrieved from a pre-stored root expansion kinetic model. The For time Functions: ; in, This represents the maximum root radius of the crop at maturity. This is the root expansion rate coefficient. The accumulated effective temperature since sowing; By the location of each crop Centered on A circular root influence domain is generated with a radius, and spatial competition segmentation of the root influence domains of adjacent plants is performed based on the Voronoi diagram to obtain the actual effective water absorption area of ​​each crop; the area of ​​the effective water absorption area is taken as the effective irrigation area of ​​a single plant. .

[0011] In conjunction with the first aspect above, in one possible implementation, the method for identifying the suspected lesion area includes: The raw RGB images captured by field cameras or inspection drones are input into the dual-branch model, which outputs the first feature map. Second feature map ; Through a learnable channel-space joint attention module and Perform weighted fusion to generate enhanced fusion feature maps. The attention weights are dynamically adjusted based on local texture complexity and global category prior. Will Input to a lightweight decoder, output pixel-level lesion probability heatmap ; Based on current light conditions and crop canopy coverage, the confidence threshold for lesion determination is dynamically calculated. The calculation formula is as follows: ; in, This is the baseline threshold under standard illumination; Given the current ambient light intensity, Ideal imaging illumination intensity; The preset adjustment coefficient; The heat map satisfies Connected regions are marked as suspected lesion areas.

[0012] In conjunction with the first aspect above, in one possible implementation, the dual-branch model includes a high-resolution shallow branch for capturing lesion edge details and a low-resolution deep branch for extracting semantic context features.

[0013] In conjunction with the first aspect above, in one possible implementation, the infection probability heatmap includes: Receive images of suspected lesion areas uploaded by an edge device, the images containing one or more marked suspected lesion areas; The images of the suspected lesion areas are preprocessed, including but not limited to size adjustment, color space conversion, and noise removal; The pre-processed image is input into a pre-trained deep learning model, which outputs the probability value of each pixel belonging to the disease infection area, forming an infection probability heatmap. The deep learning model is based on a convolutional neural network and has been trained using a large number of images labeled with different types and severity of crop diseases.

[0014] In conjunction with the first aspect above, in one possible implementation, the method for planning the optimal flight path of the drug-dispensing drone includes: Heatmap of infection probability Spatially align with the digital twin map of farmland to construct a three-dimensional operational scenario model that integrates geographic information, obstacle distribution, and crop row direction; Based on the infection probability heatmap, the target prevention and control area is determined. And assign weights to each pixel in the target prevention and control area. , This is a risk sensitivity index; among which, The preset infection determination threshold; Based on the aforementioned digital twin map of farmland and target prevention and control areas Based on the UAV's flight performance parameters, a flight path planning algorithm is used to generate a complete flight path to be optimized. Where i is the waypoint number, i = 1, 2, ..., N, each The flight path is a three-dimensional space waypoint, and includes all control points required for entering the work area, obstacle avoidance and detour, and return. Traverse the flight path Each waypoint If its horizontal projection is located at If the environmental conditions for pesticide application are met, then the waypoint is marked as a valid spraying point and added to the set of valid spraying points. ;in, Environmental application conditions are a comprehensive set of constraints for the safe and effective execution of spraying operations under the current meteorological and farmland conditions. The path planning problem is modeled as a weighted minimum flight path problem, with the objective function being: ; Constraints: ; in, From waypoint Fly to waypoint The cost; Effective spray radius; To ensure effective spraying points Centered on, effective spray radius The effective coverage area is defined by its radius. The preset weighted coverage threshold; An improved RRT algorithm combined with a greedy coverage strategy is used to solve the objective function, thereby obtaining the optimal flight path of the pesticide application drone.

[0015] In conjunction with the first aspect above, in one possible implementation, the method for obtaining the spray radius includes: For each effective spraying point The effective spray radius is dynamically calculated based on the current wind speed, wind direction, and drone flight path. : ; in, The nominal spray width, The maximum permissible wind speed for the equipment. The wind attenuation coefficient, For drones The heading angle at that location.

[0016] Secondly, this application provides an IoT-based intelligent management system for agricultural planting, including an acquisition module, an irrigation module, a disease area identification module, and a pesticide application module. The acquisition module acquires environmental data. The irrigation module dynamically calculates the current irrigation demand based on the environmental data, combined with a modified Penman-Monteith evapotranspiration model and the Crop Water Stress Index (CWSI). Based on the irrigation demand, it controls an irrigation execution unit to perform variable irrigation. The disease area identification module performs initial screening of field images for local pests and diseases through an edge computing gateway, identifies suspected lesion areas, and obtains images of these suspected lesion areas. It generates an infection probability heatmap based on the suspected lesion area images. The pesticide application module plans the optimal flight path of a pesticide application drone based on the infection probability heatmap and a digital twin map of the farmland, and controls the drone to perform targeted pesticide application.

[0017] This application provides an IoT-based intelligent management method and system for agricultural planting, which can effectively solve core problems in traditional agriculture such as extensive water, fertilizer, and pesticide management, delayed response, and serious resource waste. The system deploys a multi-source sensor network in the field to acquire key environmental data in real time, including soil moisture, air temperature and humidity, light intensity, wind speed, rainfall, and crop canopy temperature. It integrates a modified Penman-Monteith evapotranspiration model with the Crop Water Stress Index (CWSI) to dynamically calculate irrigation amounts that reflect the actual physiological needs of crops, achieving on-demand, variable, and precise irrigation. Simultaneously, it utilizes an edge computing gateway to perform initial pest and disease screening locally, uploading only images of marked suspected lesion areas, reducing bandwidth consumption and ensuring data privacy. A heat map of infection probability is then generated using a deep learning model. This heat map is further deeply integrated with a digital twin map of the farmland containing geographic information, obstacle distribution, and crop row direction to intelligently plan the optimal flight path for pesticide application drones and perform targeted spraying. The entire solution integrates the "perception-analysis-decision-execution" chain, breaking the limitations of isolated operation of irrigation and plant protection modules. For the first time, it achieves precise, coordinated water and pesticide management driven by individual plant units, physiological state, and spatial distribution. Compared to existing technologies, it significantly improves operational efficiency and control effectiveness, reduces environmental pollution, and is suitable for large-scale smart farms, providing reliable and scalable technical support for green and high-quality agricultural development.

[0018] It should be understood that the descriptions of technical features, technical solutions, beneficial effects, or similar language in this application do not imply that all features and advantages can be achieved in any single embodiment. Rather, it is understood that the description of a feature or beneficial effect means that a specific technical feature, technical solution, or beneficial effect is included in at least one embodiment. Therefore, the descriptions of technical features, technical solutions, or beneficial effects in this specification do not necessarily refer to the same embodiment. Furthermore, the technical features, technical solutions, and beneficial effects described in this embodiment can be combined in any suitable manner. Those skilled in the art will understand that embodiments can be implemented without one or more specific technical features, technical solutions, or beneficial effects of a particular embodiment. In other embodiments, additional technical features and beneficial effects may be identified in specific embodiments that do not embody all embodiments. Attached Figure Description

[0019] Figure 1 A flowchart illustrating an intelligent agricultural planting management method based on the Internet of Things (IoT) provided in this application embodiment; Figure 2 A flowchart illustrating a method for analyzing the irrigation requirements of plants provided in an embodiment of this application; Figure 3 This is a flowchart illustrating a lesion area identification method provided in an embodiment of this application. Detailed Implementation

[0020] The technical solution of the present invention will be clearly and completely described below with reference to the embodiments. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0021] The IoT-based intelligent management method for agricultural planting provided in this application embodiment can be applied to an IoT-based intelligent management system for agricultural planting, which includes: an acquisition module, an irrigation module, a disease area identification module, and a pesticide application module. The acquisition module is used to acquire environmental data; The irrigation module is used to dynamically calculate the current irrigation demand based on environmental data, combined with the modified Penman-Monteith evapotranspiration model and the crop water stress index (CWSI); and to control the irrigation execution unit to perform variable irrigation according to the irrigation demand. The disease area identification module is used to perform initial screening and identification of local pests and diseases in field images through an edge computing gateway, identify suspected lesion areas, and obtain images of suspected lesion areas; and generate an infection probability heat map based on the images of suspected lesion areas. The pesticide application module is used to plan the optimal flight path of the pesticide application drone based on the infection probability heat map and the digital twin map of the farmland, and to control the pesticide application drone to perform targeted pesticide application.

[0022] To address the technical problem that existing technologies fail to fully integrate crop physiological states and individual spatial distribution information, thus hindering precise and coordinated management of water, fertilizer, and pesticides at the individual plant scale, resulting in resource waste and low control efficiency, this application provides an IoT-based intelligent agricultural planting management method. This method includes: acquiring environmental data, including soil moisture, air temperature and humidity, light intensity, wind speed, rainfall, and crop canopy temperature; dynamically calculating the current irrigation demand based on the environmental data, combined with a modified Penman-Monteith evapotranspiration model and the crop water stress index (CWSI); controlling the irrigation execution unit to perform variable irrigation according to the irrigation demand; performing initial screening and identification of local pests and diseases on field images through an edge computing gateway, identifying suspected lesion areas, and obtaining images of suspected lesion areas; generating an infection probability heatmap based on the suspected lesion area images; and [the method is described in the original text, but the translation is incomplete]. The image contains one or more marked suspected lesion areas. Based on the infection probability heat map and the farmland digital twin map, the optimal flight path of the spraying drone is planned, and the spraying drone is controlled to perform targeted spraying. Based on this, this application collects multi-dimensional environmental data such as soil moisture, meteorological parameters, and crop canopy temperature in real time through a field sensor network. Combined with the modified Penman-Monteith model and the Crop Water Stress Index (CWSI), the actual water requirement of the crop is dynamically assessed, avoiding over- or under-watering caused by the "meter-based water supply" in traditional irrigation. At the same time, the inspection images are initially screened for diseases at the edge, and only the marked suspected lesion areas are uploaded, reducing the data transmission burden and generating a high-precision infection probability heat map. This heat map is aligned with the farmland digital twin map that integrates geographic information, obstacles, and crop row direction to plan an obstacle-avoiding and highly efficient drone spraying path to achieve targeted spraying. This solution breaks down the current disconnect between irrigation and plant protection decision-making. Based on crop physiological responses and spatial distribution, it constructs a closed loop for precision operations at the single-plant level, significantly improving resource utilization efficiency, reducing environmental impact, and providing an efficient, green, and feasible technical path for large-scale smart agriculture.

[0023] like Figure 1 As shown in the embodiment of this application, an intelligent agricultural planting management method based on the Internet of Things is provided, including: S201. Real-time collection of environmental data through an Internet of Things (IoT) sensor network deployed in farmland.

[0024] The environmental data includes soil moisture, air temperature and humidity, light intensity, wind speed, rainfall, and crop canopy temperature; the IoT sensor network consists of multiple low-power wireless sensor nodes, each of which uploads the collected data to the edge computing gateway via LoRa, NB-IoT, or Zigbee communication protocols; the environmental data specifically includes: soil volumetric water content measured by frequency domain reflectance (FDR) sensors buried in the crop root zone. The system integrates air temperature and relative humidity sensors in a miniature weather station installed 1.5 meters above the canopy; photosynthetically active radiation (PAR) or total irradiance obtained from a silicon photovoltaic light intensity sensor; wind speed measured by an ultrasonic or three-cup anemometer; rainfall recorded by a tipping bucket or capacitive rain gauge; and crop canopy temperature measured non-contactly by an infrared thermal imaging sensor during the period of maximum daily solar altitude angle (typically 12:00–14:00). ).

[0025] In some implementations, wireless sensor nodes are solar-powered and IP68 protected, supporting long-term maintenance-free operation in the field; soil moisture sensors are deployed in layers according to the depth of the main root system of the crop (e.g., 0–20cm, 20–40cm) to obtain vertical profile moisture information; infrared thermal imaging sensors are mounted on liftable brackets, and canopy temperature measurement accuracy is improved through autofocus and shadow suppression algorithms; a time synchronization mechanism (e.g., IEEE1588 PTP protocol) is used between the micro weather station and the edge gateway to ensure that the time alignment error of multi-source environmental data is less than 100 milliseconds, providing highly consistent input for subsequent evapotranspiration model and stress index calculation.

[0026] It should be noted that the aforementioned environmental data are not used in isolation, but rather serve as multimodal inputs to jointly drive the intelligent irrigation and disease early warning model: for example, air temperature and humidity are used to calculate vapor pressure difference (VPD), and combined with canopy temperature to derive the crop water stress index (CWSI); wind speed and rainfall are used to dynamically adjust the application window and spray pattern parameters; light intensity and soil moisture are used to collaboratively correct the crop coefficient. Therefore, this IoT sensor network is not only the data acquisition front end, but also the core infrastructure for realizing the "perception-decision-execution" closed loop. Its deployment density, calibration accuracy, and communication reliability directly determine the upper limit of the efficiency of the entire agricultural planting intelligent management system.

[0027] S202. Based on environmental data, combined with the modified Penman-Monteith evapotranspiration model and the crop water stress index (CWSI), the current irrigation demand is dynamically calculated; according to the irrigation demand, the irrigation execution unit is controlled to perform variable irrigation.

[0028] S203. Use an edge computing gateway to perform initial screening and identification of local pests and diseases in field images, identify suspected lesion areas, and obtain images of suspected lesion areas; generate an infection probability heat map based on the images of suspected lesion areas.

[0029] The suspected lesion area image contains one or more marked suspected lesion areas.

[0030] S204. Based on the infection probability heat map and the digital twin map of farmland, plan the optimal flight path of the spraying drone and control the spraying drone to perform targeted spraying.

[0031] Among them, the farmland digital twin map refers to a high-fidelity virtual mapping model constructed in digital space and synchronized with the physical farmland in real time. It not only contains static geographic information, but also integrates dynamic crop growth status and environmental perception data to support precision pesticide application path planning and operation simulation. The map is stored in the form of two-dimensional raster or three-dimensional point cloud, with a spatial resolution of 0.1–0.5 meters / pixel, and is strictly aligned with the actual farmland through a unified coordinate system (such as WGS-84 or local projected coordinates).

[0032] In some implementations, digital twin maps of farmland are constructed through the following steps: S1: Use drones equipped with multispectral cameras or LiDAR to regularly take aerial photographs of the target farmland to obtain high-resolution orthophotos and terrain point cloud data; S2: Based on image segmentation algorithms (such as U-Net or Mask R-CNN), identify and vectorize field boundaries, ditches, roads, utility poles, trees and other fixed obstacles to form a static geographic layer; S3: Map the real-time environmental data (including soil moisture, canopy temperature, and NDVI vegetation index) collected by the Internet of Things sensor network to the corresponding raster cells according to their spatial location to construct a dynamic attribute layer; S4: Combining planting management records (such as sowing date, variety, row spacing, and plant spacing) and crop growth models (such as GDD cumulative model), the current crop height, canopy coverage, and root distribution range are extrapolated to generate a crop physiological state layer; S5: The static geographic layer, dynamic environmental layer and crop physiological layer mentioned above are spatiotemporally fused on the edge server or cloud to form a multi-dimensional, updatable digital twin map of farmland, which is then made available for the path planning module to call via API.

[0033] It should be noted that the farmland digital twin map is not a static base map, but a "living map" that evolves continuously over time: on the one hand, its static elements (such as obstacles) are updated quarterly to adapt to farmland modifications; on the other hand, its dynamic elements (such as diseased areas and soil moisture) can be refreshed hourly or even minutely, ensuring that the pesticide application path planning is always based on the latest field conditions. In this invention, after pixel-level registration of the map with the infection probability heatmap, the spatial boundaries, shape complexity, and relative positions to obstacles of the target control area can be accurately defined, thereby supporting the generation of obstacle-avoiding, efficient, and comprehensive targeted flight paths by pesticide application drones. In addition, the digital twin map can also be used for pre-task simulation verification (such as checking whether the path crosses high-voltage lines) and post-task effect review (such as comparing changes in lesions before and after pesticide application), significantly improving the safety and closed-loop control capabilities of agricultural intelligent equipment systems.

[0034] Based on the above technical solutions, this application provides an IoT-based intelligent agricultural planting management method that addresses the resource waste and inefficient control caused by the separation of irrigation and plant protection decisions and the neglect of crop physiological states and individual spatial distribution in existing technologies. This solution achieves precise management at the individual plant scale by deeply integrating multi-source sensing and intelligent collaborative control. The system dynamically calculates the actual water requirement based on environmental data such as soil moisture and canopy temperature, combined with a modified Penman-Monteith model and the Crop Water Stress Index (CWSI), avoiding ineffective irrigation. Simultaneously, it utilizes an edge computing gateway to perform initial disease screening locally, uploading only images of suspected lesion areas, reducing bandwidth pressure and protecting privacy, and generating a pixel-level infection probability heatmap. Furthermore, this heatmap is aligned with a farmland digital twin map that integrates geographic information, obstacles, and crop row directions to plan the optimal flight path covering high-risk areas and ensuring obstacle avoidance, driving the spraying drone to perform targeted spraying. As a result, water, fertilizer, and pesticide management has been upgraded from "uniform field management" to "plant-by-plant and demand-based management," significantly improving resource utilization, enhancing early warning and precise control capabilities, and effectively solving core defects in traditional systems such as excessive irrigation, blind pesticide application, and extensive operations. This provides an efficient, green, and feasible technical path for smart agriculture.

[0035] In one possible implementation of the embodiments of this application, such as Figure 2 As shown, the above S202 can be specifically implemented through the following S301, S302 and S303, which are explained in detail below: S301. Based on soil moisture, air temperature, wind speed, and light intensity, the reference crop evapotranspiration is calculated using the FAO Penman-Monteith formula. .

[0036] Among them, reference crop evapotranspiration This refers to the evapotranspiration rate of a hypothetical reference grass surface with a height of 0.12m, a surface reflectivity of 0.23, and an aerodynamic drag of 70s / m under fully watered conditions, expressed in millimeters per day (mm / d). The FAOPenman-Monteith formula used in the calculation is as follows: ; in, The air temperature (°C) is collected in real time by a temperature sensor deployed 1.5m above the canopy. The wind speed (m / s) at a height of 2 meters is obtained by measuring a wind speed sensor and correcting for a logarithmic wind profile. The net surface radiation (MJ / m² / day) is estimated by combining the total solar radiation measured by the light intensity sensor with an empirical model. The difference between the saturated water vapor pressure and the actual water vapor pressure (kPa) is calculated from data from an air temperature and humidity sensor. The slope of the saturated water vapor pressure curve (kPa / ℃) is given. The constant of the wet and dry surface is (kPa / ℃). This represents the soil heat flux (MJ / m² / day), which is typically approximated as 0 on a diurnal scale.

[0037] In some implementations, environmental data is collected by an agricultural IoT sensor network at 5–15 minute intervals and preprocessed via an edge computing gateway. This preprocessing includes outlier removal (e.g., temperature spikes >5℃ / min), time alignment (synchronizing sensor clocks to UTC±100ms), and unit conversion (e.g., converting light intensity W / m² to solar radiation MJ / m² / day). Simultaneously, for missing net radiation... The data is interpolated using an alternative model based on sunshine hours or clear sky index; if the wind speed is measured at a non-standard height (e.g., 1m), it is corrected to a height of 2m using the logarithmic wind speed profile formula to meet the FAO standard input requirements.

[0038] It should be noted that although soil moisture is listed as one of the environmental data, it is not directly involved in the standard FAO Penman-Monteith formula. Calculations are made because they reflect the actual water supply capacity of the soil. Defined as "theoretical evapotranspiration under conditions of adequate water supply"; however, soil moisture here serves as a correction factor for the crop coefficient. Alternatively, calculate the Crop Water Stress Index (CWSI) to adjust the theoretical evapotranspiration. Converted into actual crop water requirements Therefore, mentioning soil moisture alongside meteorological parameters is intended to emphasize its synergistic value in the overall irrigation decision-making chain, rather than obscuring its role. The role in the formula.

[0039] S302, based on crop canopy temperature Using a pre-existing meteorological-canopy temperature boundary model, the theoretical minimum canopy temperature under no water stress is dynamically determined. Theoretical maximum canopy temperature under severe stress And calculate the crop water stress index. : .

[0040] in, The theoretical minimum canopy temperature (°C) under conditions of sufficient water supply and no water stress is dynamically estimated from environmental meteorological parameters (such as air temperature, humidity, and wind speed) and a pre-stored "wet edge" model. The theoretical maximum canopy temperature (°C) under severe water stress and transpiration limitation is calculated using the same "dry edge" model combined with the principle of atmospheric energy balance. The value range is [0,1]. The closer the value is to 0, the more water is available. The closer the value is to 1, the more severe the stress is.

[0041] In some implementations, the meteorological-canopy temperature boundary model includes a wet-edge model and a dry-edge model; The wet edge model assumes the crop canopy is completely wet, with unrestricted transpiration. In this state, the canopy temperature is close to the equilibrium temperature between air temperature and saturated vapor pressure difference. This model is a simplified version of the Penman-Monteith equation, with air temperature as the input parameter. Relative humidity (RH) and wind speed and solar radiation Output ; Dry edge model: This model assumes that crop transpiration is inhibited (e.g., due to root water shortage), and the canopy temperature rises to an upper limit determined by net environmental radiation and soil heat flux. This model is typically established using empirical regression or energy closure methods, and the input parameters include... , , and Output .

[0042] It should be pointed out that, although It can be measured directly, but and It is not a fixed constant, but a dynamic variable that depends on the weather conditions of the day. Therefore, relying solely on a single... The value alone cannot determine the state of water stress; it must be calculated in real time in conjunction with current weather parameters. and Only then can the CWSI be accurately obtained.

[0043] S303. Obtain the corresponding crop coefficient based on the current crop growth stage. In conjunction with the effective irrigation area per plant Compared with the preset stress sensitivity coefficient The current irrigation demand is dynamically calculated using the following formula. : .

[0044] Among them, irrigation water demand This refers to the amount of water supplied per unit time or unit area to meet crop transpiration requirements and alleviate water stress; its unit is usually mm / d or L / plant. This formula incorporates crop coefficients. , crop evapotranspiration Effective irrigation area per plant and dynamic stress response mechanism This constitutes a nonlinear, adaptive precision irrigation decision model.

[0045] It should be noted that this formula is established from the perspective of precision irrigation decision-making driven by the coupling of crop physiological response and environmental stress, aiming to achieve dynamic irrigation quantity calculation based on real-time growth status and water stress level. Its core meaning is: current irrigation demand. It depends not only on the crop's evapotranspiration potential ( Furthermore, by introducing the crop water stress index... The nonlinear inhibition term reflects the decline in the actual water demand capacity of crops under adverse conditions such as drought or high temperature. Among them, This represents the crop coefficient that varies with the growth stage. For reference evaporation, The effective irrigation area per plant. , where is the stress sensitivity coefficient, controlling the steepness of the stress response; when (Without coercion) This indicates that water is supplied according to the maximum water demand; when hour, The rate of decrease decreases as stress intensifies, reflecting the intelligent regulation concept of "supply determined by demand and disaster mitigation." This model integrates agronomic mechanisms and data-driven features to form an adaptive nonlinear irrigation decision-making framework suitable for high-precision variable irrigation systems.

[0046] Based on the above technical solutions, traditional irrigation decisions often rely on fixed growth period table lookups or estimations based on single meteorological parameters, which fail to reflect the actual water requirements of crops under real environmental stress, easily leading to "over-irrigation waste" or "water shortage and yield reduction." Especially under adverse conditions such as drought and high temperatures, crop transpiration is inhibited; if water is still supplied according to the theoretical maximum water requirement, it will result in ineffective water consumption. Conversely, ignoring stress signals may delay the timing of water replenishment. To solve this problem, this technical solution integrates the FAO Penman-Monteith formula to calculate reference evapotranspiration. The crop water stress index (CWSI) is dynamically calculated by combining canopy temperature and boundary layer models, and a nonlinear response term is introduced. This method achieves adaptive water demand regulation. Based on physiological mechanisms, it uses canopy temperature as a direct observation indicator to accurately quantify the degree of water stress; through... Reflecting differences in growth stages, combined with the effective irrigation area per plant To achieve individualized water supply; the final construction The model retains the scientific basis of the standard evapotranspiration model while enhancing its responsiveness to environmental changes. Its advantages lie in enabling a leap from "experience-based irrigation" to "physiologically driven" irrigation, reducing water waste, and simultaneously ensuring crop yield and quality. It is suitable for high-precision variable irrigation systems in smart agriculture.

[0047] In one possible implementation of this application embodiment, the crop coefficient The methods for obtaining the data include the following S401 to S404: S401. Real-time acquisition of visible and near-infrared reflectance data of crop canopy through multispectral imaging sensors deployed in farmland, calculation of normalized vegetation index (NDVI) to reflect canopy greenness and leaf area index (LAI), the higher the value, the denser the canopy and the stronger the background interference. The cumulative effective accumulated temperature (GDD) output by the soil-crop coupling model is acquired synchronously. The GDD is obtained by dynamic integration of the daily average air temperature and the crop baseline development temperature. S402, Construct a two-dimensional growth phase feature vector by fusing NDVI and GDD. .

[0048] The Normalized Difference Vegetation Index (NDVI) is calculated by real-time acquisition of reflectance data of crop canopy in the red (approximately 650 nm) and near-infrared (approximately 850 nm) bands using multispectral imaging sensors deployed in farmland. Its expression is: ; in, and These represent the reflectance in the near-infrared and red light bands, respectively; the NDVI value range is... In healthy green vegetation, the value is typically between 0.2 and 0.9, reflecting the canopy leaf area index, greenness, and photosynthetic activity.

[0049] Accumulated effective temperature (GDD) is calculated using a soil-crop coupling model based on daily average air temperature. With crop-specific baseline development temperature The result is obtained through dynamic integration, and the calculation formula is as follows: GDD characterizes the heat accumulation required for a crop to complete a certain developmental stage, and is irreversible and variety-specific.

[0050] The two-dimensional feature vector formed by combining NDVI and GDD It also encodes the crop's current physiological state (spatial dimension) and development process (temporal dimension), providing highly discriminative input for subsequent accurate identification of detailed growth stages.

[0051] In some implementations, multispectral imaging sensors are mounted on fixed field poles, mobile robots, or low-altitude drone platforms, with sampling frequencies ranging from hourly to daily. Data consistency is improved through radiometric calibration and atmospheric correction. GDD calculations rely on daily maximum / minimum temperatures collected by micro-weather stations deployed within the field, which are then arithmetically averaged and integrated. NDVI and GDD are spatiotemporally aligned using geographic coordinates and timestamps to ensure they originate from the same crop plot and on the same date. Feature vectors... After being standardized to the [0,1] interval, the data is input into a lightweight classifier to eliminate dimensional differences and improve the model's convergence speed.

[0052] It should be noted that using NDVI or GDD alone can lead to misjudgment of growth stages. For example, under drought stress, crop NDVI may decrease significantly due to leaf wilting, but GDD may have already reached the flowering threshold. In this case, relying solely on NDVI will misjudge the crop as being in the vegetative growth stage. Conversely, in cold years, GDD accumulates slowly, and even if NDVI shows a lush canopy, the crop may not have entered the reproductive stage. Therefore, the fusion of NDVI and GDD is not a simple splicing, but rather a complementary mechanism to eliminate environmental interference from a single indicator, significantly improving the robustness of growth stage identification. Furthermore, the design of this two-dimensional feature vector balances computational efficiency and information density, avoiding the introduction of too many redundant bands (such as EVI and SAVI), making it suitable for deployment on resource-constrained edge devices.

[0053] S403, The feature vector The data is input into a pre-trained lightweight crop growth stage classifier. The growth stage classifier performs transfer learning on a large number of labeled field samples and outputs the label of the current sub-growth stage. The sub-growth stages include, but are not limited to: seedling stage, tillering / branching stage, jointing / budding stage, flowering stage, grain filling / bulging stage, and maturity stage. The classifier is based on the MobileNetV3 architecture with attention mechanism enhancement.

[0054] It should be noted that the generalization ability of lightweight classifiers relies on high-quality labeled data. Although transfer learning can alleviate the small sample size problem, if the training data only covers specific varieties, densities, or climate zones (such as only winter wheat in North China), "domain shift" may occur when generalizing to southern rice or cotton in arid regions, leading to identification errors in the subdivision stage. Therefore, the classifier needs to be fine-tuned periodically by introducing new local samples through an online learning mechanism.

[0055] S404. Based on the detailed growth stage labels, dynamically retrieve the corresponding benchmark crop coefficients from the cloud-based crop knowledge graph. The final crop coefficient is obtained by adaptively correcting it based on current environmental stress factors. : ; Among them, environmental stress factors include the current air-water vapor pressure difference. and measured root zone soil volumetric water content ; The optimal VPD threshold for this growth stage; Field holding capacity; Points of wilting in the field; This represents the environmental sensitivity weighting coefficient.

[0056] Based on the above technical solutions, the traditional crop coefficient Obtaining data based on fixed growth periods using lookup tables or empirical models is difficult to adapt to the actual growth conditions of different varieties, climate zones, and field management, leading to biased irrigation decisions. Especially in complex farmland environments, crop growth is dynamically influenced by multiple factors such as light, temperature, and humidity. Single time indicators (such as calendar dates) cannot accurately reflect the true developmental process, easily resulting in "early" or "late" water supply. To address this issue, this technical solution uses multispectral imaging sensors deployed in farmland to collect canopy visible light and near-infrared reflectance in real time, calculates the Normalized Difference Vegetation Index (NDVI), and combines this with the accumulated effective temperature (GDD) output from the soil-crop coupling model to construct a two-dimensional growth phase feature vector. The data is input into a lightweight MobileNetV3 classifier enhanced with an attention mechanism to achieve high-precision identification of crop growth stages. This method integrates spatial biomass information and temporal thermodynamics, overcoming the limitations of traditional single-factor judgment; furthermore, it combines a dynamic retrieval benchmark from a cloud-based crop knowledge graph. Furthermore, it incorporates current environmental stress factors (such as VPD and soil moisture content) for adaptive correction, ultimately obtaining accurate results. Its advantage lies in achieving a leap from "static table lookup" to "dynamic perception," significantly improving... The estimation accuracy supports on-demand water supply for variable irrigation systems, improves water resource utilization efficiency, reduces production costs, and is suitable for large-scale smart agriculture scenarios.

[0057] In one possible implementation, the effective irrigated area per plant The methods for obtaining the data include the following S501 to S504: S501 uses multi-view RGB-D cameras deployed in the field or lidar mounted on drones to perform three-dimensional point cloud scanning of the target crop area and generate a high-precision three-dimensional canopy model.

[0058] In some implementations, multi-view RGB-D cameras are installed at a fixed height (e.g., 2–3 meters) on field gantry frames or mobile platforms, and multi-angle point cloud fusion is achieved through synchronous triggering and timestamp alignment; or a drone equipped with a lightweight lidar (e.g., Livox Mid-40) flies along a preset route to collect three-dimensional point cloud data at a frequency of 5–10 Hz, and combines RTK-GNSS / IMU pose information for motion compensation and georegistration, ultimately generating a three-dimensional model of the crop canopy with a resolution of centimeters.

[0059] S502. Based on the 3D model, the improved DBSCAN clustering algorithm is used to separate the spatial coordinates of individual crops. .

[0060] In some implementations, the improved DBSCAN clustering algorithm introduces height constraints and density adaptation mechanisms: the point cloud is vertically layered (e.g., 0–0.5m for the stem base, 0.5–1.5m for the canopy), clustering is performed within each layer, and the neighborhood radius and minimum number of points are dynamically adjusted using prior plant dimensions (e.g., row spacing, plant spacing); the cluster centers are calculated using the weighted centroid method, and the spatial coordinates of individual plants are output. A k-nearest neighbor graph (k=4 or 6) is constructed based on the coordinates of all plants, and Euclidean distance is used as the edge weight to form a plant spatial adjacency graph that reflects the actual planting pattern in the field, which is used for subsequent root zone division and irrigation area allocation.

[0061] S503. Based on the crop type and current growth stage, retrieve the horizontal root expansion radius from the pre-stored root expansion dynamics model. , For time Functions: ; in, This represents the maximum root radius of the crop at maturity. This is the root expansion rate coefficient. This refers to the accumulated effective temperature since sowing.

[0062] In some implementations, the root horizontal expansion kinetic model obtains root distribution data at different growth stages through field root excavation experiments and three-dimensional scanning (such as X-ray CT or ground-penetrating radar), and constructs an empirical regression model by combining it with accumulated effective temperature (GDD); among which, the parameters... Set according to crop type (e.g., corn maturity period) Wheat is ),coefficient Obtained by fitting actual observation data using the least squares method, the typical value range is [value range missing]. The model uses time... As the independent variable, output the root spread radius that dynamically changes with the growth process. It also supports daily updates to drive the spatiotemporal evolution of individual plant irrigation areas.

[0063] S504, by the position of each crop Centered on A circular root influence domain is generated with a radius, and spatial competition segmentation of the root influence domains of adjacent plants is performed based on the Voronoi diagram to obtain the actual effective water absorption area of ​​each crop; the area of ​​the effective water absorption area is taken as the effective irrigation area of ​​a single plant. .

[0064] In some implementations, the location coordinates of each crop are used. Centered on the dynamic root system expansion radius An initial circular root influence domain is generated with radius 1. A two-dimensional Voronoi diagram is constructed based on the spatial distribution of all plants, where each Voronoi cell represents the theoretical competitive region of that plant without prior constraints. Next, a Boolean intersection operation is performed on each circular root influence domain and its corresponding Voronoi cell to achieve fair division of overlapping areas. Specifically, when the root influence domains of two crops overlap, their common portion is divided along the perpendicular bisector (Voronoi edge), ensuring that each plant is allocated only its "dedicated" underground space. The resulting polygonal region is the actual effective water absorption area of ​​the plant, and its area is accurately calculated using the Shoelace formula or a GIS engine and recorded as the effective irrigation area per plant. This is used for the individualized allocation of subsequent variable irrigation amounts.

[0065] It should be noted that Voronoi diagram segmentation assumes that soil resources are spatially uniformly distributed and that competition among plants is symmetrical. This idealized premise may be broken in actual farmland due to soil heterogeneity (such as local compaction or fertility patches) or differences in plant growth (such as diseased versus healthy plants). For example, a dominant plant may "encroach" on the water of a weaker neighbor through stronger root activity, causing Voronoi segmentation to overestimate the effective water absorption area of ​​the weaker plant. Therefore, in high-precision applications, a weighted Voronoi diagram (e.g., weighted based on NDVI or canopy volume) can be introduced to shift the segmentation boundary toward the weaker plant, thus more accurately reflecting the competition pattern of underground resources. Furthermore, this method assumes that the horizontal root extension is an isotropic circle. If a row-direction guiding effect exists (such as in ridge cultivation), an elliptical influence domain should be used instead of the circle, and then intersection with the Voronoi cell should be performed to improve model fidelity.

[0066] Based on the above technical solutions, traditional irrigation systems mostly adopt the "irrigation per acre" or "irrigation per row" model, ignoring the spatial competition and water absorption differences among individual crop roots, resulting in uneven water resource distribution and low fertilizer and water utilization rates. Especially in densely planted farmland, the root systems of adjacent plants overlap, and if their spatial distribution characteristics are ignored, it is easy to cause some plants to be "over-irrigated" while others are "under-irrigated." To solve this problem, this technical solution uses multi-view RGB-D cameras or UAV LiDAR deployed in the field to perform three-dimensional point cloud scanning of the target area, constructing a high-precision three-dimensional canopy model; and uses an improved DBSCAN clustering algorithm to separate the spatial coordinates of each crop. The time function was dynamically calculated using a GDD-driven root expansion dynamics model. This method reflects the growth pattern of the root system as it progresses through the reproductive process. A circular root influence domain is generated centered on each plant, and the overlapping areas of adjacent plants are fairly segmented using a Voronoi diagram to obtain the actual effective water absorption area for each plant. Starting from "individualized root zone modeling," this method integrates three-dimensional perception, growth models, and spatial competition mechanisms, achieving a leap from "group average" to "precise individual plant analysis."

[0067] In one possible implementation of the embodiments of this application, such as Figure 3 As shown, the above S203 can be specifically implemented through the following S601-S605, which are explained in detail below: S601. Deploy a lightweight dual-branch convolutional neural network model in an edge computing gateway.

[0068] The dual-branch model includes a high-resolution shallow branch for capturing lesion edge details and a low-resolution deep branch for extracting semantic context features.

[0069] S602. Input the raw RGB images collected by field cameras or inspection drones into the dual-branch model, and output the first feature map respectively. Second feature map ; Through a learnable channel-space joint attention module and Perform weighted fusion to generate enhanced fusion feature maps. The attention weights are dynamically adjusted based on local texture complexity and global category prior. Will Input to a lightweight decoder, output pixel-level lesion probability heatmap .

[0070] Among them, the attention weights are dynamically adjusted based on local texture complexity and global category prior; in the process of the dual-branch model processing the original RGB image, the detail feature map ( ) and semantic feature maps ( The feature maps capture both fine structural information and high-level semantic content of the image. To effectively fuse these two types of feature maps, a learnable channel-space joint attention module is used. and Perform weighted fusion to generate an enhanced fusion feature map. In this process, the attention weights are dynamically adjusted based on the texture complexity of the local region and the class prior knowledge of the entire image. Specifically, for regions with complex textures and rich information, the mechanism assigns higher attention weights to preserve this key information more meticulously; while for parts with clear class attributes, the relevant features are strengthened through global class priors to ensure that the final fused feature map has both sufficient detail and important semantic information.

[0071] In some implementations, various approaches may be used to optimize the above process in different application scenarios. For example, in specific crop disease identification tasks, specialized attention mechanisms can be designed for different types of crops or diseases to improve detection accuracy and robustness. Furthermore, transfer learning methods can be combined, utilizing the knowledge gained by pre-trained models on large-scale datasets to initialize model parameters, thereby accelerating the training process and improving model performance. Simultaneously, employing multi-scale feature fusion strategies is also a common practice, integrating feature information at different resolutions to enable the model to better handle lesion manifestations at various scales.

[0072] S603. Based on current light conditions and crop canopy coverage, dynamically calculate the confidence threshold for lesion determination. The calculation formula is as follows: ; in, This is the baseline threshold under standard illumination, which is usually obtained through training with a large number of labeled samples; The current ambient light intensity is obtained in real time by light sensors deployed in the fields or by spectrometers carried by drones; The ideal imaging illumination intensity represents the standard value for optimal imaging quality; Preset adjustment coefficients are used to control the degree of influence of light attenuation and canopy shading on the threshold, which are usually calibrated through cross-validation.

[0073] The core of this formula is that when there is insufficient light ( The reduced image contrast makes lesion detection more difficult, so the lesion determination confidence threshold needs to be lowered to avoid missed detections. When the canopy is too dense (high NDVI), lesions are easily obscured and the signal is weak, so the lesion determination confidence threshold should also be appropriately lowered to enhance sensitivity.

[0074] In some implementations, light intensity The light is collected once per minute by a silicon photovoltaic cell-type light sensor installed on a field pole, and then input into the model after atmospheric scattering correction; NDVI is calculated from multispectral images simultaneously captured by the same drone while performing disease inspection tasks, ensuring spatiotemporal consistency. The initial values ​​were obtained by training on disease identification datasets of major crops such as corn and rice, and stored in a cloud-based crop knowledge graph; the adjustment coefficients... , Automatic matching based on crop type—for example, for wheat with thin leaves that are easily affected by shade, A larger value (0.3) is chosen to enhance the response to canopy shading; however, for cotton with thick leaves, A smaller value (0.1) can be taken; the final calculated dynamic threshold It is sent to the edge inference device in real time for subsequent binarization processing.

[0075] The heat map satisfies Connected regions are marked as suspected lesion areas.

[0076] In some implementations, the heatmap satisfies... The pixels are binarized to generate a binary mask image, where foreground pixels (value 1) represent high-probability lesion areas. A neighborhood-based connected component labeling algorithm is used to identify all interconnected foreground regions and assign a unique label to each connected region. To further improve the discrimination accuracy, morphological post-processing operations are applied to each connected region, including opening (removing isolated noise) and closing (filling internal holes). At the same time, false positive regions that do not conform to typical lesion morphology are filtered out by combining preset geometric constraints (such as minimum area threshold and maximum aspect ratio). The finally retained connected regions are marked as "suspected lesion regions" and their boundary contours, center coordinates and average confidence scores are output for subsequent targeted application path planning or agronomist review.

[0077] Based on the above technical solutions, traditional crop disease identification methods rely on single visual features or fixed threshold segmentation, which struggles to cope with complex lighting conditions, shading, and background interference in the field, resulting in high false positive rates and poor robustness. Especially in cloudy, backlit, or densely canopied scenes, the color difference between lesions and healthy tissue decreases, making it easy to miss lesions relying solely on semantic information; while relying solely on texture details may misidentify non-disease areas such as leaf damage and insect infestation as lesions. To address this issue, this technical solution employs a dual-branch model to extract detailed features from the image separately. (e.g., edges, spots) and semantic features (e.g., overall lesion morphology), and dynamically weighted fusion is performed through a learnable channel-space joint attention module to generate an enhanced fusion feature map. This effectively balances local texture and global semantic information. The lightweight decoder outputs pixel-level lesion probability heatmaps. And introduce a system based on light intensity Dynamic threshold of NDVI The calculation formula allows the judgment threshold to adaptively adjust with environmental changes: the threshold is lowered to improve sensitivity when there is insufficient light, and noise is further suppressed when NDVI is high (dense canopy). This scheme achieves a leap from "static threshold" to "dynamic perception," significantly improving the accuracy and generalization ability of disease detection. It is suitable for intelligent plant protection systems in variable farmland environments, supporting precision application and early warning.

[0078] S604. Receive an image of a suspected lesion area uploaded by an edge device, the image containing one or more marked suspected lesion areas; and preprocess the suspected lesion area image, including but not limited to size adjustment, color space conversion and noise removal.

[0079] S605. Input the pre-processed image into the pre-trained deep learning model. The deep learning model outputs the probability value of each pixel belonging to the disease infection area, forming an infection probability heatmap.

[0080] The deep learning model is built upon convolutional neural networks and has been trained using a large number of images labeled with different crop disease types and their severity. The deep learning model is a fully convolutional semantic segmentation network, with its backbone architecture based on improved U-Net, DeepLabv3+, or HRNet convolutional neural network structures, specifically designed for pixel-level disease region identification. During the training phase, the model used a large-scale dataset collected in the field and meticulously labeled by plant protection experts. Each image has a pixel-level label indicating the crop species, disease type (e.g., rice blast, maize leaf spot, wheat stripe rust), and severity level (e.g., mild, moderate, severe). Through end-to-end learning, the model can automatically extract multi-scale texture, color, edge, and contextual semantic features from the input images and output an infection probability heatmap with the same spatial dimensions as the input image. , where each pixel value This indicates the confidence level that the location belongs to the disease-infected area.

[0081] In some implementations, preprocessing includes: illumination normalization of the original RGB image (e.g., Retinex algorithm), dehazing enhancement (for rainy or foggy scenes), color space conversion (e.g., converting to HSV or Lab to enhance the contrast between lesions and healthy tissue), and uniform cropping or padding to the model input requirements (e.g., 512×512); the deep learning model is deployed on edge computing devices (e.g., Jetson AGX Orin) or cloud inference servers, using TensorRT or ONNXRuntime for acceleration; to improve the detection rate of small lesions, the model embeds attention mechanisms (e.g., CBAM or SE modules) and atrous convolution in the encoder-decoder structure to expand the receptive field; the output infection probability heatmap can be further fused with multispectral NDVI maps or canopy temperature maps to distinguish fungal lesions from mechanical damage or insect traces.

[0082] It should be noted that although this deep learning model can generate high-resolution heatmaps of infection probability, its performance is highly dependent on the diversity and labeling quality of the training data. If the training set lacks samples of a certain disease under specific light, growth stage, or variety, the model may miss or misjudge in practical applications. In addition, the heatmaps output by the model usually contain some noise (such as blurred edges and isolated false positives), and post-processing strategies (such as Conditional Random Fields (CRF), morphological filtering, or dynamic thresholding) are required to generate reliable suspected lesion areas. More importantly, this model only reflects "visible lesions" and cannot identify latent infections. Therefore, it should be used in conjunction with physiological sensor data (such as chlorophyll fluorescence and volatile organic compounds (VOCs)) to build a multimodal early warning system.

[0083] Furthermore, the infection probability heatmap is corrected based on crop type, growth stage, and environmental factors to obtain the final infection probability heatmap.

[0084] It should be noted that the initial correction of the infection probability heatmap includes: Using crop growth models to predict the expected appearance characteristics of crops under their current health status; The probability values ​​in the infection probability heatmap are adjusted by combining meteorological data, soil moisture, and temperature information to reflect the impact of actual environmental conditions on disease development. The final output is an infection probability heatmap, which not only indicates the presence or absence of the disease, but also quantifies the likelihood of infection in each area, and provides users with an intuitive visualization interface to assist in decision-making.

[0085] Based on the above technical solutions, traditional crop disease identification relies on manual field inspections or fixed-threshold image processing methods, which suffer from low efficiency, strong subjectivity, and high misjudgment rates, making it difficult to meet the needs of rapid diagnosis in large-scale farmland. Especially in complex field environments, factors such as changes in light intensity, leaf shading, and the indistinctness of early lesions make it difficult for conventional methods to accurately identify disease types and severity. To address this issue, this technical solution receives images of suspected lesion areas uploaded by edge devices and combines them with a deep learning model to achieve automated, high-precision intelligent disease identification. The images are preprocessed (e.g., size normalization, color space conversion, noise removal) to improve input quality; then input to a deep learning model built on a convolutional neural network (CNN). This model has been thoroughly trained on a large dataset of images labeled with different crop disease types and severity levels, possessing powerful feature extraction and classification capabilities. The model outputs the probability value of each pixel belonging to a disease-infected area, generating an infection probability heatmap, achieving pixel-level lesion localization and quantitative assessment. This solution shifts from "experience-based judgment" to "data-driven" approaches, significantly improving identification speed and accuracy. It supports concurrent identification and dynamic grading of multiple diseases and is applicable to scenarios such as drone inspections and intelligent agricultural machinery. It enables early detection, early warning, and precise pesticide application, reducing pesticide usage and improving the level of intelligent agricultural production.

[0086] In one possible implementation of this application embodiment, the above-mentioned S204 can be specifically implemented by the following S701, S702 and S703, which are described in detail below: S701, Infection probability heatmap Spatially align with the digital twin map of farmland to construct a three-dimensional operational scenario model that integrates geographic information, obstacle distribution, and crop row direction.

[0087] In some implementations, infection probability heatmaps Spatial alignment with the farmland digital twin map is achieved through the following steps: Using UAV RTK-GNSS / IMU pose data recorded synchronously during heat map acquisition, the heat map is transformed from an image coordinate system to a unified geographic coordinate system (such as WGS-84 or local projection coordinates); based on the vectorized field boundaries, crop row lines, and fixed obstacles (such as utility poles, ditches, and trees) layers in the farmland digital twin map, the heat map undergoes affine transformation and resampling to ensure its spatial resolution (typically 0.1–0.5 m / pixel) matches that of the digital twin. The map is strictly matched; the aligned heat map is overlaid as an attribute layer onto the 3D scene of the digital twin map; the ground elevation is provided by LiDAR point cloud or DEM model, the crop row orientation information is used to constrain the directionality of subsequent spraying paths, and the obstacle distribution is embedded in the scene in the form of 3D voxels or grids; the final 3D operation scene model not only includes the spatial distribution of disease risk, but also integrates accessibility, operation safety boundary and agronomic structure information, providing a high-fidelity environmental perception foundation for obstacle avoidance trajectory planning and targeted spraying control of spraying drones.

[0088] S702. Determine the target prevention and control area based on the infection probability heat map. And assign weights to each pixel in the target prevention and control area. , This is a risk sensitivity index; among which, This is the preset infection determination threshold.

[0089] In some implementation methods, the target prevention and control area Determination of infection risk weight function The construction is achieved through the following process: the infection probability heatmap output by the deep learning model is... Input is sent to the edge computing unit; based on the current crop type, disease type, and growth stage, the corresponding infection determination threshold is dynamically loaded from the cloud-based crop knowledge graph. (For example, rice sheath blight occurs during the tillering stage) =0.45, while corn rust takes place during the tasseling stage. =0.6); perform binarization operation to satisfy The pixels are marked as areas requiring prevention and control, forming an initial target area mask. Based on this, the mask is morphologically optimized using agronomic expert rules (such as removing isolated patches smaller than 100 pixels and filling internal holes), ultimately generating a connected and structurally sound mask. Simultaneously, the system automatically configures risk sensitivity indices based on prevention and control strategies. And calculate the weighting function in real time. Among them, areas with high infection probability receive higher weight due to the power function amplification effect, and are thus prioritized for coverage in subsequent path planning.

[0090] S703, based on farmland digital twin map, target prevention and control area Based on the UAV's flight performance parameters, a flight path planning algorithm is used to generate a complete flight path to be optimized. .

[0091] Where i is the waypoint number, i = 1, 2, ..., N, each The flight path consists of three-dimensional waypoints and includes all control points required for entering the work area, obstacle avoidance, and return.

[0092] S704, Traversing Flight Path Each waypoint If its horizontal projection is located at If the environmental conditions for pesticide application are met, then the waypoint is marked as a valid spraying point and added to the set of valid spraying points. ;in, Environmental application conditions are a comprehensive set of constraints for the safe and effective execution of spraying operations under the current meteorological and agricultural conditions.

[0093] In some implementations, flight path Generated by RRT or A algorithm, the complete flight path includes entry into the work area, obstacle avoidance and detour, and return; each waypoint Includes three-dimensional coordinates and heading angle The system obtains real-time wind speed, temperature, and humidity data by calling the local meteorological API, and performs spatial queries by combining the obstacle layer in the farmland digital twin map; for each waypoint, its horizontal projection is first determined. Whether it falls Within the scope (achieved through point-polygon inclusion detection), it is then checked whether the environmental application conditions are met; if both are met, it is added to the set of valid spraying points. ;in, This indicates that the actual number of waypoints available for spraying is less than or equal to the total number of waypoints; this process can be completed on edge computing devices with a millisecond-level response, supporting dynamic task adjustment.

[0094] S705. The path planning problem is modeled as a weighted minimum flight path problem, with the objective function being: ; Constraints: ; in, From waypoint Fly to waypoint The cost; Effective spray radius; To ensure effective spraying points Centered on, effective spray radius The effective coverage area is defined by its radius. This is the preset weighted coverage threshold.

[0095] It should be noted that the core idea of ​​this formula is to satisfy the condition that "the target prevention and control area is effectively covered at least..." Under the premise of "proportional weighted risk area", find a route from waypoints The optimal flight path This minimizes the sum of travel costs (such as flight distance, energy consumption, or time) between adjacent waypoints on the path. The cost function is... Characterizing from waypoints Flying towards The actual cost; the constraint requires all effective spray points. With spray radius The covered area must cover the target area. Total weighted risk of infection in China The efficiency is more than doubled, ensuring that the prevention and control effect is not sacrificed. Therefore, this model strikes a balance between "efficiency" and "effectiveness": it pursues both low energy consumption and short flight range economy, and ensures full coverage of high-risk lesion areas, which is the core optimization framework for achieving precise and intelligent pesticide application.

[0096] The method for obtaining the effective spray radius is as follows: for each effective spray point... The effective spray radius is dynamically calculated based on the current wind speed, wind direction, and drone flight path. : ; in, The nominal spray width, The maximum permissible wind speed for the equipment. The wind attenuation coefficient, For drones The heading angle at that location.

[0097] It should be noted that the formula for calculating the effective spray radius is established from the perspective of dynamic modeling of spray coverage under the coupled influence of environmental disturbances and flight attitude, aiming to achieve real-time adaptive adjustment of the effective spray radius in UAV pesticide application operations. Its core meaning is: for each effective spray point... According to the current wind speed ,wind direction and the drone's heading angle Dynamically calculate its actual effective spray radius that can be covered. .in, The nominal spray width is under windless conditions; This indicates the percentage of spray attenuation caused by wind speed. The higher the wind speed, the stronger the spray diffusion and the smaller the effective coverage area. This reflects the influence of the angle between the wind direction and the flight direction. When the flight direction is consistent with the wind direction (tailwind), the spray is blown further, resulting in enhanced effective coverage; when facing the wind, the spray flows back, limiting coverage. Therefore, this formula comprehensively considers three major factors: wind force, wind direction angle, and flight direction, achieving refined modeling of spraying performance. This ensures effective control coverage is maintained even under complex weather conditions, improving application accuracy and resource utilization.

[0098] S706. An improved RRT algorithm combined with a greedy coverage strategy is used to solve the objective function and obtain the optimal flight path of the pesticide application drone.

[0099] Based on the above technical solutions, traditional agricultural drone operations mostly employ fixed flight paths or simple grid coverage patterns, making it difficult to address the uneven distribution of diseases and complex farmland terrain. This leads to the problem of "overspraying disease-free areas and underspraying severely diseased areas," resulting in pesticide waste and reduced control effectiveness. To solve this problem, this technical solution uses an infection probability heatmap... Spatially aligned with a digital twin map of farmland, a 3D operational scenario model is constructed that integrates geographic information, obstacle distribution, and crop row direction, accurately identifying high-risk areas and assigning them weights. This enables "on-demand medication." Based on this, flight path planning is modeled as a weighted minimum-cost path optimization problem, aiming to minimize the total flight cost while simultaneously achieving a preset threshold for weighted coverage. The method considers not only path length and energy consumption but also ensures sufficient coverage of high-risk areas. An improved RRT algorithm combined with a greedy coverage strategy is used to solve the problem, efficiently generating the optimal trajectory while ensuring obstacle avoidance safety. This scheme achieves a leap from "uniform spraying" to "targeted precision," significantly improving application efficiency and resource utilization, reducing environmental impact, and is suitable for intelligent plant protection systems in complex farmland environments, promoting the green, efficient, and intelligent development of agriculture.

[0100] Some of the data in the above formula are calculated by removing dimensions and taking their numerical values. The formula is the closest to the real situation obtained by software simulation of a large amount of collected data. The preset parameters and preset thresholds in the formula are set by those skilled in the art according to the actual situation or obtained through simulation of a large amount of data.

Claims

1. An Internet of Things-based intelligent management method for agricultural planting, characterized in that, include: Acquire environmental data, including soil moisture, air temperature and humidity, light intensity, wind speed, rainfall, and crop canopy temperature; Based on the environmental data, and combining the modified Penman-Monteith evapotranspiration model and the Crop Water Stress Index (CWSI), the current irrigation demand is dynamically calculated; and the irrigation execution unit is controlled to perform variable irrigation according to the irrigation demand. The edge computing gateway performs local pest and disease screening on field images to identify suspected lesion areas and obtain images of suspected lesion areas; an infection probability heat map is generated based on the images of suspected lesion areas; the images of suspected lesion areas contain one or more marked suspected lesion areas; Based on the infection probability heatmap and the farmland digital twin map, the optimal flight path of the spraying drone is planned, and the spraying drone is controlled to perform targeted spraying. 2.The Internet of Things-based intelligent management method for agricultural planting according to claim 1, characterized in that, The method for obtaining the current irrigation demand includes: Based on the aforementioned soil moisture, air temperature, wind speed, and light intensity, the reference crop evapotranspiration was calculated using the FAO Penman-Monteith formula. ; Based on the crop canopy temperature Using a pre-existing meteorological-canopy temperature boundary model, the theoretical minimum canopy temperature under no water stress was determined. Theoretical maximum canopy temperature under severe stress And calculate the crop water stress index. : ； Obtain the corresponding crop coefficient based on the current crop growth stage. In conjunction with the effective irrigation area per plant Compared with the preset stress sensitivity coefficient Calculate the current irrigation demand. : 。 3. The intelligent agricultural planting management method based on the Internet of Things according to claim 2, characterized in that, The crop coefficient is obtained based on the current crop growth stage. ,include: The normalized vegetation index (NDVI) is calculated by collecting real-time visible and near-infrared reflectance data of crop canopy using multispectral imaging sensors deployed in farmland. The cumulative effective accumulated temperature (GDD) output by the soil-crop coupling model is acquired synchronously. The GDD is obtained by dynamic integration of the daily average air temperature and the crop baseline development temperature. The NDVI and GDD are fused to construct a two-dimensional growth phase feature vector. ; The feature vector The data is input into a pre-trained lightweight crop growth stage classifier, which outputs the label of the current sub-growth stage. The classifier is based on the MobileNetV3 architecture with enhanced attention mechanism. The sub-growth stages include, but are not limited to: seedling stage, tillering / branching stage, jointing / budding stage, flowering stage, grain filling / enlargement stage, and maturity stage. Based on the detailed growth stage labels, the corresponding benchmark crop coefficients are dynamically retrieved from the cloud-based crop knowledge graph. The final crop coefficient is obtained by adaptively correcting it based on current environmental stress factors. : ； Among them, environmental stress factors include the current air-water vapor pressure difference. and measured root zone soil volumetric water content ; The optimal VPD threshold for this growth stage; Field holding capacity; Points of wilting in the field; This represents the environmental sensitivity weighting coefficient.

4. The intelligent agricultural planting management method based on the Internet of Things according to claim 2, characterized in that, The effective irrigation area per plant The methods for obtaining it include: By using multi-view RGB-D cameras deployed in the field or lidar mounted on drones, a three-dimensional point cloud scan of the target crop area is performed to generate a three-dimensional model of the canopy. Based on the aforementioned three-dimensional canopy model, the improved DBSCAN clustering algorithm was used to separate the spatial coordinates of individual crop plants. ; Based on the crop type and current growth stage, the horizontal root expansion radius is retrieved from a pre-stored root expansion kinetic model. The For time Functions: ； in, This represents the maximum root radius of the crop at maturity. This is the root expansion rate coefficient. The accumulated effective temperature since sowing; By the location of each crop Centered on A circular root influence domain is generated with a radius, and spatial competition segmentation of the root influence domains of adjacent plants is performed based on the Voronoi diagram to obtain the effective irrigation area for each crop. .

5. The intelligent agricultural planting management method based on the Internet of Things according to claim 1, characterized in that, The method for identifying the suspected lesion area includes: The raw RGB images captured by field cameras or inspection drones are input into the dual-branch model, which outputs the first feature map. Second feature map ; Through a learnable channel-space joint attention module and Perform weighted fusion to generate enhanced fusion feature maps. The attention weights are dynamically adjusted based on local texture complexity and global category prior. Will Input to a lightweight decoder, output pixel-level lesion probability heatmap ; Based on current light conditions and crop canopy coverage, the confidence threshold for lesion determination is dynamically calculated. The calculation formula is as follows: ； in, This is the baseline threshold under standard illumination; Given the current ambient light intensity, Ideal imaging illumination intensity; Normalized Difference Vegetation Index; The preset adjustment coefficient; The heat map satisfies Connected regions are marked as suspected lesion areas.

6. The intelligent agricultural planting management method based on the Internet of Things according to claim 5, characterized in that, The dual-branch model includes a high-resolution shallow branch for capturing lesion edge details and a low-resolution deep branch for extracting semantic context features.

7. The intelligent agricultural planting management method based on the Internet of Things according to claim 1, characterized in that, The infection probability heatmap includes: Receive images of suspected lesion areas uploaded by an edge device, the images containing one or more marked suspected lesion areas; The images of the suspected lesion areas are preprocessed, including but not limited to size adjustment, color space conversion, and noise removal; The preprocessed image is input into a pre-trained deep learning model, which outputs the probability value of each pixel belonging to the disease infection area, forming an infection probability heatmap. The deep learning model is based on a convolutional neural network and has been trained using a large number of images labeled with different crop disease types and their severity.

8. The intelligent agricultural planting management method based on the Internet of Things according to claim 1, characterized in that, The method for planning the optimal flight path of the drug-dispensing drone includes: Heatmap of infection probability Spatially align with the digital twin map of farmland to construct a three-dimensional operational scenario model that integrates geographic information, obstacle distribution, and crop row direction; Based on the infection probability heatmap, the target prevention and control area is determined. And assign weights to each pixel in the target prevention and control area. , This is a risk sensitivity index; among which, The preset infection determination threshold; Based on the aforementioned digital twin map of farmland and target prevention and control areas Based on the UAV's flight performance parameters, a flight path planning algorithm is used to generate a complete flight path to be optimized. Where i is the waypoint number, i = 1, 2, ..., N, each The flight path is a three-dimensional space waypoint, and includes all control points required for entering the work area, obstacle avoidance and detour, and return. Traverse the flight path Each waypoint If its horizontal projection is located at If the environmental conditions for pesticide application are met, then the waypoint is marked as a valid spraying point and added to the set of valid spraying points. ;in, Environmental application conditions are a comprehensive set of constraints for the safe and effective execution of spraying operations under the current meteorological and farmland conditions. The path planning problem is modeled as a weighted minimum flight path problem, with the objective function being: ； Constraints: ; in, From waypoint Fly to waypoint The cost; Effective spray radius; To ensure effective spraying points Centered on, effective spray radius The effective coverage area is defined by its radius. This is the preset weighted coverage threshold; An improved RRT algorithm combined with a greedy coverage strategy is used to solve the objective function, thereby obtaining the optimal flight path of the pesticide application drone.

9. The intelligent agricultural planting management method based on the Internet of Things according to claim 8, characterized in that, The method for obtaining the spray radius includes: For each effective spraying point The effective spray radius is dynamically calculated based on the current wind speed, wind direction, and drone flight path. : ； in, The nominal spray width, The maximum permissible wind speed for the equipment. The wind attenuation coefficient, For drones The heading angle at that location.

10. An intelligent management system for agricultural planting based on the Internet of Things (IoT), operating based on the intelligent management method for agricultural planting based on the IoT as described in any one of claims 1-9, characterized in that, It includes an acquisition module, an irrigation module, a disease area identification module, and a pesticide application module; The acquisition module is used to acquire environmental data; The irrigation module is used to dynamically calculate the current irrigation demand based on the environmental data, combined with the modified Penman-Monteith evapotranspiration model and the Crop Water Stress Index (CWSI); and to control the irrigation execution unit to perform variable irrigation according to the irrigation demand. The disease area identification module is used to perform initial screening of local pests and diseases in field images through an edge computing gateway, identify suspected lesion areas, and obtain images of suspected lesion areas; and generate an infection probability heat map based on the images of suspected lesion areas. The application module is used to plan the optimal flight path of the application drone based on the infection probability heat map and the farmland digital twin map, and to control the application drone to perform targeted application.

Images