Plant protection unmanned aerial vehicle pesticide application path planning method and system based on deep learning

Abstract

The invention discloses a plant protection unmanned aerial vehicle pesticide application path planning method and system based on deep learning, and relates to the technical field of smart agriculture, and the method comprises the steps: constructing meta-training task distribution of a virtual farmland scene in an offline stage, carrying out the architecture design and parameter optimization through a meta-learning framework, and obtaining a path planning agent; the architecture of the path planning agent is configured to extract time sequence invariant features from environment observation, generate a composite action instruction, and train the path planning agent by adopting a meta-reinforcement learning algorithm; deploying the trained path planning intelligent agent to the plant protection unmanned aerial vehicle, executing initial exploration flight on a target new farmland, and collecting a small amount of interaction data; and performing optimization and updating through feature extraction parameters of the path planning intelligent agent. According to the invention, collaborative improvement of plant protection unmanned aerial vehicle path planning in efficiency, adaptability and agricultural long-term benefits is realized.

Description

Technical Field

[0001] This invention relates to the field of smart agriculture technology, and in particular to a method and system for planning pesticide application paths for plant protection drones based on deep learning. Background Technology

[0002] With the development of precision agriculture technology, agricultural drone spraying has become an important part of modern agriculture. Its operational efficiency and spraying effectiveness highly depend on the level of intelligence in flight path planning. In recent years, artificial intelligence technologies, represented by deep reinforcement learning, have been applied in this field. Existing solutions typically involve constructing a simulation environment and collecting a large amount of data to train an end-to-end deep neural network strategy. This strategy can directly output flight control commands based on the real-time environmental information perceived by the drone, achieving a certain degree of automated obstacle avoidance and path coverage optimization, and demonstrating good performance in scenarios with sufficient training data coverage.

[0003] The limitations of existing technical solutions are that their performance is heavily dependent on the consistency between training data and the real operating environment. When a trained model is deployed to a farmland with a completely new terrain, crop type, or pest and disease pattern, the planning performance of the model often degrades due to the difference in the distribution of environmental data. In order to restore the accuracy of the operation, traditional fine-tuning methods are prone to overfitting a small amount of new data when adapting to the new environment, which leads to the loss of general decision-making knowledge learned from the original massive data, thereby affecting the stable performance of the model in different scenarios. Summary of the Invention

[0004] In view of the aforementioned existing problems, the present invention is proposed.

[0005] Therefore, this invention provides a deep learning-based method for pesticide application path planning of agricultural drones, which solves the problem of insufficient generalization ability of existing technologies in novel scenarios and the balance between knowledge forgetting and overfitting during the adaptation process.

[0006] To solve the above-mentioned technical problems, the present invention provides the following technical solution: In a first aspect, the present invention provides a method for pesticide application path planning of agricultural drones based on deep learning, which includes, in the offline stage, constructing a meta-training task distribution of a virtual farmland scene, and performing architecture design and parameter optimization through a meta-learning framework to obtain a path planning agent; The architecture of the path planning agent is configured to extract time-invariant features from environmental observations, generate composite action instructions, and train the path planning agent using a meta-reinforcement learning algorithm. The trained path planning agent was deployed on an agricultural drone and used to conduct an initial exploratory flight over the target new farmland to collect a small amount of interactive data. The path planning agent is optimized and updated by extracting features from its parameters, and the decision-making correlations generated during the update process are used to guide the fine-tuning of the path planning agent. Using a finely tuned path planning agent, generate and execute paths and pesticide application instructions that are both adaptable to the current farmland and optimized for future operations.

[0007] As a preferred embodiment of the deep learning-based agricultural drone spraying path planning method of the present invention, the following steps are included in the offline stage: Constructing a meta-training task distribution for a virtual farmland scene includes the following steps: Based on the meta-training task distribution dataset, a dynamic growth model is generated for each virtual farmland scene to simulate the temporal evolution of the virtual farmland scene's state over multiple consecutive operation cycles. Based on geographic information data of real farmland, crop planting models and historical agricultural data, a virtual farmland scene description is generated through parametric modeling and data augmentation techniques. By integrating temporal evolution data of state with virtual farmland scene descriptions, a meta-training task distribution is formed.

[0008] As a preferred embodiment of the deep learning-based agricultural drone spraying path planning method of the present invention, the path planning agent is obtained through architecture design and parameter optimization using a meta-learning framework, including the following steps: Based on the meta-training task distribution, the network structure parameters of the path planning agent are initialized, and the inner and outer loop optimization processes of the meta-learning framework are executed using the meta-training task distribution. The network structure parameters of the path planning agent are updated through the inner and outer loop optimization process of the meta-learning framework. Based on the updated network structure parameters, the path planning agent is obtained.

[0009] As a preferred embodiment of the deep learning-based agricultural drone spraying path planning method of the present invention, the path planning agent is configured to extract time-invariant features from environmental observations and generate composite action commands, including the following steps: Configure the architecture of the path planning agent as a feature encoder component and a policy decoder component; The feature encoder component receives farmland images and environmental observation data of UAV flight status, and processes the environmental observation data to extract time-invariant feature representations. The policy decoder component outputs composite action instructions based on time-invariant feature representation.

[0010] As a preferred embodiment of the deep learning-based pesticide application path planning method for agricultural drones described in this invention, the path planning agent is trained using a meta-reinforcement learning algorithm, including the following steps: Virtual farmland scenes are sampled from the meta-training task distribution. In the simulation environment of the current meta-task, the path planning agent interacts with the environment to generate interaction trajectory data. Based on the dynamic growth model and interaction trajectory data corresponding to the virtual farmland scene, the cumulative reward of time series reward is calculated; Based on the cumulative reward of interactive trajectory data and time-series rewards, the parameters of the path planning agent are updated by executing a meta-reinforcement learning algorithm. After sampling and training a preset number of virtual farmland scenarios in the meta-training task distribution, the trained path planning agent is obtained.

[0011] As a preferred embodiment of the deep learning-based pesticide application path planning method for agricultural drones described in this invention, the method involves: deploying a trained path planning agent onto an agricultural drone and performing an initial exploratory flight over a new target farmland to collect a small amount of interactive data, including the following steps: The trained path planning agent is loaded into the onboard computer memory of the agricultural drone, and the agricultural drone is launched to fly to the airspace of the target new farmland. The plant protection drone was controlled to fly over the target new farmland in a preset initial exploration flight mode; During the initial exploration flight, the path planning agent generates tentative action commands, records the environmental observation data acquired by the sensors during the initial exploration flight, the tentative action commands generated by the path planning agent, and the corresponding real-time performance evaluation data, forming a small amount of interactive data.

[0012] As a preferred embodiment of the deep learning-based agricultural drone spraying path planning method of the present invention, the method includes the following steps: Optimization and updating of feature extraction parameters of the path planning agent: Environmental observation data from a small amount of interactive data is input into the feature encoder component of the path planning agent to obtain feature representations. These representations are then input into a fixed policy decoder component to generate predicted actions. The loss between the predicted actions and the actual recorded actions is calculated. The backpropagation algorithm is used to calculate the parameter gradient of the loss with respect to the feature encoder component of the path planning agent. Based on the parameter gradient, multiple sets of candidate parameter update vectors for the feature encoder component are generated. Multiple sets of candidate parameter update vectors are applied to the feature encoder component to generate multiple temporary updated versions of the feature encoder component. Multiple temporary updated versions of the feature encoder component are used to process environmental observation data in a small amount of interactive data to obtain multiple sets of intermediate feature representations. Multiple sets of intermediate feature representations are input into the policy decoder component of the path planning agent to generate multiple sets of candidate action instruction sequences. The decision consistency measure of the multiple sets of candidate action instruction sequences is evaluated, and the candidate parameter update vector with the highest decision consistency measure is selected to optimize and update the parameters of the feature encoder component of the path planning agent.

[0013] As a preferred embodiment of the deep learning-based agricultural drone spraying path planning method of the present invention, the following steps are included: Guiding the path planning agent with decision-making correlations generated during the update process, and fine-tuning the agent: The distribution differences among multiple sets of candidate action instruction sequences are analyzed to identify decision correlation patterns. The guidance direction for the policy decoder component of the path planning agent is determined based on the decision correlation pattern; Following the guiding direction of the policy decoder component of the path planning agent, the weight mapping relationship of the policy decoder component to the intermediate feature representation is adjusted. By combining the optimized feature encoder component parameters with the adjusted policy decoder component weight mapping relationship, the path planning agent is fine-tuned.

[0014] As a preferred embodiment of the deep learning-based agricultural drone spraying path planning method of the present invention, the method includes: using a fine-tuned path planning agent to generate and execute spraying instructions that are both adaptable to the current farmland and optimized for future operations, comprising the following steps: The fine-tuned path planning agent is connected to the real-time sensor data stream of the agricultural drone, and the fine-tuned path planning agent processes the current environmental observations from the sensor data stream in real time. Based on current environmental observations, the fine-tuned path planning agent generates an optimized flight path that is adaptable to the current farmland. Combining the optimized flight path with the long-term optimization goals embedded in the path planning agent, a matching precise pesticide application instruction is generated. The optimized flight path and precise application instructions are sent to the flight control unit and application unit of the agricultural drone; The flight control unit and the drug delivery unit work together to execute optimized flight paths and precise drug delivery commands.

[0015] Secondly, the present invention provides a deep learning-based agricultural drone spraying path planning system, including a construction module that, in the offline stage, constructs a meta-training task distribution of a virtual farmland scene, and performs architecture design and parameter optimization through a meta-learning framework to obtain a path planning agent. The training module is configured to extract time-invariant features from environmental observations, generate composite action instructions, and train the path planning agent using a meta-reinforcement learning algorithm. The data collection module deploys the trained path planning agent onto the agricultural drone and performs initial exploratory flights over the target new farmland to collect a small amount of interactive data. The fine-tuning module optimizes and updates the path planning agent by extracting its feature parameters and guides the decision-making correlations generated during the update process, thus fine-tuning the path planning agent. The execution module uses a fine-tuned path planning agent to generate and execute paths and pesticide application instructions that are both adaptable to the current farmland and optimized for future operations.

[0016] The beneficial effects of this invention are as follows: By constructing a meta-training task distribution that integrates a dynamic growth model, and training a path planning agent with long-term optimization awareness through meta-reinforcement learning based on this time-series distribution, the agent acquires a rapid adaptive meta-ability for new farmland. Through small-sample fine-tuning guided by decision correlation, the feature encoder and policy decoder are simultaneously optimized on a small amount of new data, which not only quickly adapts to the characteristics of new farmland, but also maintains the stability and consistency of decision logic. The agent can generate and execute optimized paths and pesticide application instructions in real time that accurately adapt to the current real-time conditions of the farmland and create favorable agronomic conditions for future operation cycles. This solves the core problems of insufficient model generalization ability, slow adaptation process, and easy loss of general knowledge faced by existing technologies in new and unusual farmland scenarios, and realizes a synergistic improvement in the efficiency, adaptability, and long-term agronomic benefits of agricultural drone path planning. Attached Figure Description

[0017] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0018] Figure 1 This is a flowchart of a deep learning-based method for pesticide application path planning using agricultural drones.

[0019] Figure 2 This is a schematic diagram of a deep learning-based agricultural drone spraying path planning system. Detailed Implementation

[0020] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0021] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and those skilled in the art can make similar extensions without departing from the spirit of the invention. Therefore, the invention is not limited to the specific embodiments disclosed below.

[0022] Secondly, the term "one embodiment" or "example" as used herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the invention. The appearance of an embodiment in different places in this specification does not necessarily refer to the same embodiment, nor is it a single or selective embodiment that mutually excludes other embodiments.

[0023] Reference Figures 1-2 This is one embodiment of the present invention, which provides a method for pesticide application path planning for agricultural drones based on deep learning, including the following steps: S1. In the offline phase, construct the meta-training task distribution of the virtual farmland scene.

[0024] S1.1 Based on the meta-training task distribution dataset, generate a dynamic growth model corresponding to each virtual farmland scene to simulate the temporal evolution data of the virtual farmland scene's state over multiple consecutive operation cycles.

[0025] Furthermore, the meta-training task distribution dataset provides the basic static parameters for each virtual farmland scenario, such as the initial spatial distribution of crop types, soil properties, and initial pest and disease stress levels. The generation of dynamic growth models relies on combining agronomic knowledge with data-driven methods. Based on partial differential equations, it takes the current farmland state and the applied pesticide action as input, and internally simulates the interaction process of crop growth, pest and disease spread, and pesticide deposition and degradation to output the predicted future farmland state. For example, it simulates how the deposition distribution of pesticide in the canopy changes the local survival probability of pests and diseases after a pesticide application, thereby affecting their spatial diffusion pattern in the next growth stage. This process is repeated over multiple consecutive virtual operation cycles to generate a sequence of farmland states that evolves from the initial state over time and with a series of simulated pesticide application interventions, i.e., state temporal evolution data.

[0026] S1.2 Based on geographic information data of real farmland, crop planting models and historical agricultural data, a virtual farmland scene description is generated through parametric modeling and data augmentation techniques.

[0027] Furthermore, geographic information data of real farmland, including but not limited to field boundary coordinates, elevation information, and slope and aspect data obtained from surveying or satellite remote sensing, forms the foundation of the spatial geometric framework of virtual farmland. Crop planting models provide typical growth morphology parameters of the target crop, such as the theoretical range of plant height, canopy size, row spacing, and plant spacing. Historical agricultural data, such as records of the location and intensity of pests and diseases in past years and soil fertility distribution maps, provide statistical priors for constructing the non-uniform attribute distribution of virtual farmland. Parametric modeling refers to using the above data sources to generate specific virtual scene instances by setting a series of control parameters. For example, a polygonal field outline is generated based on geographic information data. Crop planting model parameters are used to distribute the position and initial size of crop plants within this outline according to specific rules (such as straight lines or curves). Initial pest and disease pressure levels or crop health are assigned to different areas with reference to historical agricultural data. Data augmentation techniques further increase the diversity of the scene by randomly perturbing or combining the basic parameters.

[0028] Specifically, for example, affine transformations are applied to the outline of the field to change its shape and size; randomness is introduced into the crop planting model to simulate the fluctuations in plant position and size caused by sowing deviations; or new initial distribution maps of pests and diseases that conform to the real distribution patterns are synthesized based on the statistical characteristics of historical agricultural data. The application of parametric modeling and data augmentation techniques ensures that the generated virtual farmland scene description is both realistic and sufficiently diverse in terms of spatial layout, crop configuration, and initial agricultural conditions, thus becoming the generation of virtual farmland scene descriptions.

[0029] S1.3 Integrate state time-series evolution data with virtual farmland scene descriptions to form a meta-training task distribution.

[0030] Furthermore, the temporal dynamic information is structured and encoded into the definition of the meta-learning task, constructing a task distribution with inherent temporal correlation. The state temporal evolution data contains the possible state trajectories of each virtual farmland scene under multiple hypothetical operation cycles, while the virtual farmland scene description defines the static identity and initial conditions of the scene. A task encapsulation format is constructed so that each meta-task instance is explicitly associated with a specific virtual farmland scene description and bound to a series of possible state evolution trajectories generated by the scene's dynamic growth model as its task dynamic environment. For example, a meta-task can be instantiated as completing the pesticide application operation planning for three consecutive growth cycles in a wheat field scene with a specific initial distribution of pest and disease patches. The environmental state of each cycle is calculated in real time through the bound dynamic growth model based on the state of the previous cycle and the agent's pesticide application actions. The integration method ensures that each task in the meta-training task distribution has clear initial conditions and a dynamic transition function that follows agronomic laws.

[0031] Specifically, when training on sampling tasks, the meta-reinforcement learning framework presents the agent not with isolated, static decision fragments, but with a complete, causally continuous, multi-cycle decision sequence. This forces the agent to develop the ability to model and plan the long-term consequences of actions during the learning process, thus laying a crucial training foundation for achieving the goal of rapidly adapting to new farmland and considering long-term benefits. Through the above integration, a meta-training task distribution rich in temporal decision structures is formed for meta-learning.

[0032] S2. The path planning agent is obtained by designing the architecture and optimizing the parameters through the meta-learning framework.

[0033] S2.1. Based on the meta-training task distribution, initialize the network structure parameters of the path planning agent, and use the meta-training task distribution to execute the inner and outer loop optimization processes of the meta-learning framework.

[0034] Furthermore, a two-layer optimization paradigm is applied to enable the path planning agent to acquire rapid adaptation capabilities. The network structure parameters of the path planning agent are first assigned random or pre-trained initial values. The meta-training task distribution provides a large number of virtual farmland scene tasks with temporal dynamics. The inner loop optimization process of the meta-learning framework refers to sampling a specific virtual farmland task from the meta-training task distribution and performing several steps of gradient updates on the current path planning agent in this task environment. The goal of this update is to rapidly improve the performance of the path planning agent on specific tasks, simulating the rapid adaptation process on new tasks. The outer loop optimization process, after adapting to multiple different tasks in the inner loop, evaluates the comprehensive performance of the initially adapted path planning agent on all sampled tasks or retained validation tasks. Based on this evaluation result, the initial network structure parameters of the path planning agent are updated through backpropagation.

[0035] Specifically, the inner loop forces the path planning agent to learn how to adjust its internal representation based on limited experience to cope with new scenarios, while the outer loop optimizes the initial parameters. This means that when adapting to the inner loop from the initial point, the training mechanism enables the path planning agent to ultimately learn not a strategy specific to a particular farmland, but a general strategy with the ability to quickly adjust and generalize.

[0036] S2.2. Through the inner and outer loop optimization process of the meta-learning framework, the network structure parameters of the path planning agent are updated, and the path planning agent is obtained based on the updated network structure parameters.

[0037] Furthermore, in each round of meta-training iteration, the inner loop optimization process generates a series of adapted temporary path planning agent versions across multiple sampling tasks. The outer loop optimization process then calculates the loss function of each temporary version on its respective task, backtracks, and accumulates it onto the initial path planning agent network structure parameters, thereby performing an update to the initial parameters. For example, the outer loop update might adjust the initial parameters in such a direction that, when facing any new task in the future, only a very small modification to the inner loop parameters is needed to enable the path planning agent to output an efficient sequence of actions. This process is repeated, and the diverse virtual farmland scenarios in the meta-training task distribution ensure that the path planning agent is exposed to various possible terrains, crop layouts, and pest and disease evolution patterns.

[0038] Specifically, after multiple rounds of such iterative optimization, the initial network structure parameters of the path planning agent are updated to a state with high generalization and rapid adaptation potential, encoding a meta-knowledge or learning algorithm, making the path planning agent itself an entity with excellent initialization and efficient adaptation mechanism. Based on the updated network structure parameters, a path planning agent with strong cross-scene generalization and rapid adaptation capabilities is obtained.

[0039] The architecture of the S3 path planning agent is configured to extract time-invariant features from environmental observations and generate compound action instructions.

[0040] S3.1 Configure the architecture of the path planning agent as a feature encoder component and a policy decoder component.

[0041] Furthermore, the end-to-end decision-making process is decoupled into two functionally defined and collaborative sub-networks. The architecture of the path planning agent consists of a feature encoder component and a policy decoder component connected in series, achieving functional separation and specialized optimization of perception and decision-making. The feature encoder component is specifically responsible for processing high-dimensional, redundant raw environmental observation data containing a large amount of irrelevant details, with the goal of extracting the high-level abstract information most relevant to the decision. The policy decoder component focuses on receiving abstract features and generating specific, executable compound action instructions based on them.

[0042] Specifically, the network is forced to learn a hierarchical representation: the feature encoder component learns to filter out irrelevant noise in environmental observation data that varies with seasons, lighting, and the drone's perspective, while the policy decoder component learns to map stable high-level features to the action space. For example, the feature encoder component may learn to identify the boundaries and relative positions of healthy and pest-infested areas in the crop canopy, while the policy decoder component calculates the optimal speed and nozzle flow rate to reach the target area based on boundary information and the drone's own state. This separation allows for independent pre-training or specialized architecture design of the feature encoder component to improve perception robustness, while the policy decoder component can remain lightweight to achieve rapid decision-making. This completes the configuration of the path planning agent as a cascaded architecture of the feature encoder and policy decoder components.

[0043] S3.2 The feature encoder component receives environmental observation data of farmland images and UAV flight status. The feature encoder component processes the environmental observation data to extract time-invariant feature representations.

[0044] Furthermore, the feature encoder component is endowed with the ability to extract stable and essential features from multimodal and dynamic inputs. Environmental observation data includes high-dimensional spatial information such as farmland images and low-dimensional state vectors such as drone flight status. The feature encoder component needs to fuse and process heterogeneous data. A convolutional neural network branch is used to process farmland images to extract spatial features, while a fully connected network branch is used to process drone flight status. The features extracted by the two branches are then fused. Through training, the feature encoder component is guided to focus on those elements that are relatively stable in the time dimension and are crucial to decision-making, while ignoring interfering factors that change instantaneously or periodically.

[0045] Specifically, for example, the structure of crop rows and the shape and relative position of major pest and disease patches in farmland images are relatively stable high-level semantic features, while light conditions and the posture of leaves swaying in the wind are instantaneously changing low-level appearance features. The feature encoder component is trained through meta-reinforcement learning to extract the former as a time-invariant feature representation. This is because only stable features can provide a consistent and reliable basis for decision-making at different times and in different operating cycles. This capability enables the path planning agent to generate consistent feature representations when facing observations of the same farmland at different growth stages or under different weather conditions, thereby ensuring the stability of the decision. This gives the feature encoder component the ability to distill the core stable information required for decision-making from the original observations. The feature encoder component outputs a time-invariant high-level feature vector that integrates spatial semantics and state information, which is the time-invariant feature representation.

[0046] S3.3 The policy decoder component outputs compound action instructions based on time-invariant feature representation.

[0047] Furthermore, the policy decoder component receives time-invariant feature representations from the feature encoder component. These representations are compact encodings of key environmental information. The policy decoder component typically consists of a series of fully connected layers that decode this high-dimensional feature vector into a set of specific, coordinated action parameters, i.e., composite action commands. These commands include the UAV's three-dimensional spatial displacement increments or target waypoints, flight speed, and flow control commands for the pesticide application system. The training objective of the policy decoder component is to learn a mapping policy from state features to optimal actions. The time-invariant feature representations, as input, ensure that the information basis upon which the policy decoder component makes decisions is stable and reliable, thereby improving the consistency of the output actions when the policy faces similar farmland conditions at different time points.

[0048] Specifically, for example, based on the time-invariant characteristic that represents the presence of strip-shaped pest and disease zones in the area ahead, the strategy decoder component can decode a series of coordinated action commands, such as adjusting the flight course to be parallel to the strip, reducing the flight altitude, and increasing the pesticide application rate. By outputting composite commands, the decoupling and suboptimal problems that may occur in traditional methods, such as planning the path first and then generating speed and pesticide application commands, are avoided. This achieves integrated and coordinated decision-making for flight and pesticide application, and completes the final conversion from environmental understanding to specific control commands. The strategy decoder component outputs composite action commands containing multi-dimensional control quantities such as displacement, speed, and flow rate.

[0049] S4. The path planning agent is trained using a meta-reinforcement learning algorithm.

[0050] S4.1 Sample virtual farmland scenes from the meta-training task distribution. In the simulation environment of the current meta-task, the path planning agent interacts with the environment to generate interaction trajectory data.

[0051] Furthermore, the core component of meta-reinforcement learning training—policy evaluation and experience collection for individual tasks—is the meta-training task distribution, a collection containing a large number of diverse virtual farmland scenarios and corresponding dynamic growth models, which is repeatedly sampled during training. Each sampling yields a specific virtual farmland scenario and its dynamic growth model, constituting a current meta-task. In the simulation environment of the current meta-task, the path planning agent receives the environmental observation state generated by the simulation environment based on its current policy network parameters and outputs composite action instructions. The simulation environment updates the farmland state according to the dynamic growth model and provides immediate rewards. The continuous state-action-reward-new state interaction sequence between the path planning agent and the simulation environment is recorded, forming an interaction trajectory data that includes the agent's complete decision sequence and environmental feedback under that specific task.

[0052] Specifically, by switching between different virtual farmland scenarios, the path planning agent is able to access and learn decision-making experience in dealing with various initial conditions, terrain layouts, and pest and disease evolution patterns. The generation of interactive trajectory data simulates various operational situations that the path planning agent may experience in real farmland, generating an interactive trajectory data containing a sequence of states, actions, and rewards for the current meta-task.

[0053] S4.2 Calculate the cumulative reward of time series rewards based on the dynamic growth model and interaction trajectory data corresponding to the virtual farmland scene.

[0054] Furthermore, long-term agronomic effects are quantified and integrated into the optimization objectives of standard reinforcement learning, thereby guiding the path planning agent to make forward-looking decisions. Interaction trajectory data records the actions taken by the path planning agent at each step within a work cycle, along with the immediate operational rewards from environmental feedback. However, immediate operational rewards alone are insufficient to assess the potential impact of a single pesticide application on the future state of the farmland. Therefore, the concept of time-series rewards is introduced. Calculating time-series rewards requires utilizing a dynamic growth model corresponding to the virtual farmland scenario. The dynamic growth model can predict the long-term impact of an action on the farmland state multiple time steps later, based on the state and actions recorded in the interaction trajectory data at a specific moment.

[0055] Specifically, for example, applying pesticides to a specific area at a certain moment may yield a good immediate coverage reward, but the dynamic growth model can predict that over-application in that area may lead to pesticide damage to subsequent crops or alterations in the local ecology, thus negatively impacting future pest and disease control. This negative impact is calculated as a negative time-series reward. For each time step in the interaction trajectory data, in addition to obtaining the immediate reward given by the task reward function, a forward simulation is performed through the dynamic growth model to evaluate the cumulative impact of the current action on the future state, and this impact is quantified as a time-series reward. The comprehensive reward is the sum of the immediate task reward and the weighted time-series reward, while the cumulative reward is the discounted sum of the comprehensive rewards of all steps in the entire interaction trajectory data. This ensures that while pursuing the optimal immediate task effect, the path planning agent must intrinsically consider the long-term agronomic consequences of its decisions, thereby learning sustainable and beneficial task strategies throughout the entire growth cycle. The cumulative time-series reward value corresponding to the interaction trajectory data is calculated as a comprehensive indicator to evaluate the performance of the path planning agent on this meta-task.

[0056] The cumulative return expression is: ; in, For trajectory The cumulative return The total number of time steps for the trajectory. Index for the current time step, for, For a moment The overall reward obtained by the intelligent agent. For the homework reward function, For a moment Environmental observation status For a moment The actions performed by the intelligent agent For time-series reward weights, For time-series rewards, This is interactive trajectory data.

[0057] S4.3. Based on the cumulative reward of the interaction trajectory data and the time-series reward, the meta-reinforcement learning algorithm is executed to update the parameters of the path planning agent. After sampling and training a preset number of virtual farmland scenes in the meta-training task distribution, the trained path planning agent is obtained.

[0058] Furthermore, based on the cumulative rewards of interaction trajectory data and time-series rewards, meta-reinforcement learning algorithms (such as MAML or Reptile) execute their parameter update rules. The cumulative reward serves as a performance indicator, and its gradient information is used to adjust the network parameters of the path planning agent. Under the meta-learning framework, the update is divided into two layers: inner loop update and outer loop update. It focuses more on describing how to use single or multiple interaction trajectory data and their cumulative rewards to drive the adaptive adjustment of the path planning agent's parameters. For example, the algorithm may calculate the expected gradient of the cumulative reward after the current policy interacts on multiple sampled tasks, update the initial parameters of the path planning agent along the direction of increasing the expected cumulative reward, and the update process is repeatedly iterated. In each iteration, a batch of new virtual farmland scenes are sampled from the meta-training task distribution to generate new interaction trajectory data, obtain the cumulative reward containing time-series rewards, and update the parameters accordingly. By completing the sampling and training of a preset number of virtual farmland scenes in the meta-training task distribution, the parameters of the path planning agent have undergone a large number of diverse tasks.

[0059] Specifically, through meta-reinforcement learning optimization, the path planning agent learns not a fixed strategy for a single scenario, but a set of excellent initial parameters that can quickly adapt to new tasks, as well as meta-knowledge implicit in the parameters on how to adjust the strategy according to the new environment. The introduction of time-series rewards ensures that the meta-knowledge is not only related to rapid adaptation, but also to whether the adapted strategy has good long-term agronomic benefits. After sufficient training, the parameters of the path planning agent converge to a state, enabling it to have strong rapid adaptation potential and long-term planning ability when facing new farmland, thus obtaining a trained path planning agent.

[0060] S5. Deploy the trained path planning agent on the agricultural drone and conduct initial exploratory flights over the target new farmland to collect a small amount of interactive data.

[0061] S5.1 Load the trained path planning agent into the onboard computer memory of the agricultural drone and start the agricultural drone to fly to the airspace of the target new farmland.

[0062] Furthermore, the process of migrating the core intelligent model obtained from offline training from the development environment to the deployment of actual operational hardware involves a path planning agent that has been trained. This agent is a neural network model file containing feature encoder and policy decoder components. This file is loaded into the memory of the onboard computer of the agricultural drone via a data transmission interface, enabling the onboard computer to run the model for real-time inference. Subsequently, through ground station commands or autonomous mission planning, the agricultural drone starts from its takeoff point and flies to the airspace where the target new farmland is located.

[0063] Specifically, the system completes the spatial transfer and functional readiness of the intelligent decision-making entity from the virtual training environment to the real physical carrier, providing core algorithmic support for subsequent adaptive learning and task execution in the new environment. This ensures that the path planning agent model can run stably and with low latency on resource-constrained embedded hardware, guaranteeing the feasibility of real-time decision-making. The system also enables the deployment of the trained path planning agent on agricultural drones and its arrival in the operational airspace.

[0064] S5.2 Control the agricultural drone to fly over the target new farmland in the preset initial exploration flight mode.

[0065] Furthermore, it proactively and efficiently collects preliminary structured information about new farmland, rather than passively waiting or flying randomly. The preset initial exploration flight mode is a predefined flight path template that balances coverage and safety. For example, it could be a bow-shaped scan or a square-shaped boundary patrol covering part of the farmland. Controlling the agricultural drone to fly according to this mode means that in the initial stage, it does not rely on a path planning agent to make fully autonomous operational decisions, but instead executes a known and safe flight plan.

[0066] Specifically, it can ensure the basic safety of drones flying in completely unfamiliar environments and avoid risks that may arise from the agent's strategy not adapting to the new environment; the patterned flight can systematically acquire environmental observation data at different locations of the target new farmland, avoiding the blindness and randomness of data collection. The exploration process itself is also a rapid preliminary survey of the target new farmland, providing a controllable and efficient platform for data collection. The plant protection drone completed the initial exploration flight over the target new farmland according to the preset mode.

[0067] S5.3 During the initial exploration flight, the path planning agent generates tentative action commands, records the environmental observation data acquired by the sensors during the initial exploration flight, the tentative action commands generated by the path planning agent, and the corresponding real-time performance evaluation data, forming a small amount of interactive data.

[0068] Furthermore, in the crucial data preparation stage for achieving rapid adaptation with small samples, the online exploration process is combined with data collection and strategy trial operation. During the initial exploratory flight, the path planning agent is not idle but activated and runs in parallel. It receives environmental observation data from the agricultural drone's sensors (such as cameras, IMUs, and positioning modules) in real time and generates composite action commands based on the new data. These commands are considered exploratory action commands and are not used to directly control the drone (flight may still mainly follow a preset pattern). However, they represent the path planning agent's initial response to the new environment based on its current knowledge. The flight control unit may partially integrate the exploratory commands for fine-tuning, for example, while maintaining the general direction of following the preset path, fine-tuning the flight altitude or attitude according to the agent's commands. More importantly, the entire process is recorded synchronously: the raw environmental observation data acquired by the sensors, the exploratory action commands generated by the path planning agent, and the real-time performance evaluation data based on the current state and action estimation are all saved together.

[0069] Specifically, each exploration flight not only involves environmental perception but also serves as a closed-loop strategy test and data production. The collected data pairs form state-action-reward triplets, which are precisely the interactive data required for reinforcement learning. Due to the limited time of the exploration flight, the scale of the collected interactive data is small. However, this small amount of data contains key feedback information about the actual decision-making attempts and results of the path planning agent in the new environment. This transforms the exploration flight into a data acquisition experiment, forming a small amount of interactive data for subsequent rapid fine-tuning.

[0070] S6. Optimize and update the parameters by extracting features from the path planning agent.

[0071] S6.1 Input the environmental observation data from a small amount of interaction data into the feature encoder component of the path planning agent to obtain feature representations. Input the fixed policy decoder component to generate predicted actions. Calculate the loss between the predicted actions and the actual recorded actions. Calculate the parameter gradient of the loss with respect to the feature encoder component of the path planning agent through the backpropagation algorithm. Generate multiple sets of candidate parameter update vectors for the feature encoder component based on the parameter gradients.

[0072] Furthermore, the gradient of the loss with respect to the parameters of the feature encoder component of the path planning agent is calculated using the backpropagation algorithm. Based on the parameter gradient, multiple sets of candidate parameter update vectors for the feature encoder component are generated. This is the first step in initiating parameter optimization in small-sample fine-tuning. A supervised feature adaptation task is constructed, and on this basis, the purposeful parameter update direction is explored. Environmental observation data from a small amount of interaction data is used as input and processed by the feature encoder component under the current parameter state to obtain the corresponding feature representation. Since the policy decoder component is fixed at this stage, the predicted action generated after the feature representation is input into the policy decoder component should theoretically be as close as possible to the actual recorded action in the same set of data. The difference between the two is quantified by the action loss function. The calculated overall loss function value reflects the degree of adaptation of the current feature encoder component to the feature extraction of new farmland data. Through the backpropagation algorithm, the gradient of the overall loss with respect to each network parameter of the feature encoder component can be accurately calculated. This gradient vector indicates the direction of the steepest parameter descent to reduce prediction error.

[0073] Specifically, updating along this gradient direction may lead to overfitting or damage to feature generalization in small sample scenarios. By strategically applying different perturbations or transformations, multiple sets of candidate parameter update vectors can be generated. For example, small-amplitude random noise vectors in different directions can be superimposed on the gradient direction to explore the neighborhood, or different dimensions of the gradient vector can be scaled to different degrees to simulate the importance adjustment of different feature channels. This constructs multiple sets of candidate update schemes with subtle or significant differences in update direction and amplitude. It acknowledges that a single gradient direction may not be the optimal or most robust choice with limited data. Instead, by constructing a set of candidate update directions, the gradient of the basic parameters is obtained, and multiple sets of candidate parameter update vectors for exploration are generated.

[0074] The overall loss function is expressed as follows: ; in, For parameters The overall loss function, This represents the total number of samples in a small amount of interactive data. In the sample Predicted action vectors on For the network parameter set of the feature encoder component, In the sample The actual recorded action vectors on the data. This is the sample index.

[0075] The expression for the parameter gradient is: ; in, For differential operators, For the gradient of the parameters, This represents the change in the loss function.

[0076] S6.2 Apply multiple sets of candidate parameter update vectors to the feature encoder component to generate multiple temporary updated feature encoder component versions.

[0077] Furthermore, parallel simulations of multiple candidate schemes are performed, with each set of candidate parameter update vectors representing a different parameter adjustment path. To evaluate the potential effect of each path, they are added to the current parameters of the feature encoder component, thereby virtually creating multiple versions of the feature encoder component with different parameter states. These versions are temporary and do not permanently replace the original parameters; rather, they exist as temporary copies to evaluate the effects of different update directions. For example, the first set of candidate update vectors may primarily adjust the weights of the convolutional kernels responsible for extracting color features in the feature encoder component, while the second set of candidate update vectors may focus on adjusting the weights responsible for extracting texture features. Applying each set of vectors to the original parameters yields version A and version B of the feature encoder component adjusted in a specific direction.

[0078] Specifically, without actually submitting any updates or breaking the original model, we explored the possible changes in feature extraction results brought about by multiple parameter evolution paths. Based on multiple sets of candidate parameter update vectors, we generated multiple corresponding temporary updated feature encoder component versions.

[0079] S6.3. Use multiple temporarily updated feature encoder component versions to process environmental observation data in a small amount of interactive data to obtain multiple sets of intermediate feature representations.

[0080] Furthermore, by utilizing each temporarily updated version of the feature encoder component, the same environmental observation data in a small amount of interactive data is reprocessed. Since the internal parameters of each version of the feature encoder component are adjusted in different directions, their internal responses and abstractions to the same input data will also be different, thus outputting different feature representations. For example, for the same farmland image, a feature encoder component version that focuses on spatial structure may output a feature representation that emphasizes crop ridge geometry, while another version that focuses on spectral information may output a feature representation that emphasizes differences in vegetation health.

[0081] Specifically, the same batch of environmental observation data is processed by multiple different versions of feature encoder components and then transformed into multiple sets of intermediate feature representations with different focuses or expressions. The intermediate feature representations are the bridge connecting different parameter adjustment schemes and the final decision output. By comparing how different intermediate feature representations affect the decisions of subsequent policy decoder components, the effectiveness and robustness of different parameter update directions can be indirectly evaluated, thus transforming the same environmental observation data into multiple sets of different intermediate feature representations.

[0082] S6.4 Input multiple sets of intermediate feature representations into the policy decoder component of the path planning agent to generate multiple sets of candidate action instruction sequences, evaluate the decision consistency measure of the multiple sets of candidate action instruction sequences, and select the candidate parameter update vector with the highest decision consistency measure to perform optimization update on the parameters of the feature encoder component of the path planning agent.

[0083] Furthermore, multiple sets of intermediate feature representations are sequentially input into a fixed policy decoder component. Since the parameters of the policy decoder component remain unchanged, its output depends entirely on the input feature representations. Therefore, each set of intermediate feature representations leads the policy decoder component to generate a corresponding set of candidate action instruction sequences. These sequences represent potential job strategies generated based on different feature understanding methods. The focus of the evaluation is not the absolute performance of the candidate action instruction sequences (due to a lack of sufficient validation data), but rather the measure of decision consistency among them. The measure of decision consistency can be quantified as the similarity or concentration of multiple sets of candidate action instruction sequences in their statistical distribution. For example, the entropy value of the action sequence can be extracted; a lower entropy value indicates that the decisions generated by different parameter versions are more concentrated and consistent. Alternatively, the similarity between each pair of different action sequences can be calculated and averaged.

[0084] Specifically, with limited data, minimizing a single loss function might lead to a solution that fits the current few samples well but has poor generalization. However, if multiple different parameter update directions can result in highly consistent action decisions output by the policy decoder component, different feature representations are equivalent or convergent at a higher, more abstract decision level. Consistency is more likely to reflect a reliable capture of the essential characteristics of the new farmland rather than an overfit to noise. Selecting the candidate parameter update vector with the highest decision consistency means selecting the update path that most stably supports the original decision logic. Finally, this selected update vector is formally applied to the parameters of the feature encoder component of the path planning agent to complete this optimization update, greatly improving the robustness and generalization of small-sample fine-tuning. The optimal update direction is determined by evaluating decision consistency, and the final optimization update is performed on the parameters of the feature encoder component.

[0085] S7. Guide the path planning agent by analyzing the decision-making correlations generated during the update process.

[0086] S7.1 Analyze the distribution differences among multiple sets of candidate action instruction sequences to identify decision correlation patterns.

[0087] Furthermore, in the in-depth analysis of multiple candidate action instruction sequences, the evaluation of parameter updates shifts from a single loss value to the macroscopic statistical characteristics of multiple strategy outputs. The multiple candidate action instruction sequences come from different versions of feature encoder components processing the same data, and they constitute a set of action decisions. To analyze the distribution differences, it is first necessary to regard the action instructions in each candidate action instruction sequence as a set of points in a high-dimensional space.

[0088] Specifically, for example, it might be observed that most candidate action command sequences have very similar values ​​in a certain spatial dimension. This indicates that the policy decoder component's control decisions in this dimension are strongly correlated, i.e., highly consistent, under different feature extraction methods. Conversely, in another dimension, the values ​​of each sequence are dispersed, indicating weak correlation in that dimension and susceptibility to differences in feature extraction. Alternatively, it might be found that the overall distribution of some candidate action command sequences differs significantly from other sequences, exhibiting outliers. Identifying decision correlation patterns involves quantifying and understanding consistency, divergence, and outliers. When the parameters of the feature encoder component change in different directions, how does the fixed policy decoder component map these changes to the final action space? This allows us to find the feature change patterns corresponding to the parameter update directions that maintain stable and consistent decision output. Through statistical analysis, key decision correlation patterns were identified from multiple sets of candidate action command sequences.

[0089] S7.2 Determine the guiding direction of the policy decoder component for the path planning agent based on the decision correlation model.

[0090] Furthermore, the internal mapping of the policy decoder component is proactively adjusted using the correlation patterns obtained from the analysis, rather than simply selecting the update direction of the feature encoder component. Decision correlation patterns reveal the stable and sensitive regions of the policy decoder component's output decision under multiple sets of intermediate feature representation inputs. Based on these patterns, the guiding direction for the policy decoder component is determined. For example, if the analysis reveals that for the current new farmland data, the policy decoder component exhibits high consistency when mapping features to the flight speed dimension, but shows significant divergence when mapping to the pesticide application rate dimension, then the guiding direction might be to strengthen the policy decoder component's attention to the feature dimensions that influence the application rate decision, or to adjust its internal weights to reduce the sensitivity of the application rate output to certain noisy dimensions in the features.

[0091] Specifically, the guiding direction can be translated into suggestions for adjusting the connection weights of the policy decoder component, or a bias correction of its activation function. The process of determining the guiding direction essentially uses multi-version feature extraction results as probes to diagnose and understand the behavior of the fixed policy decoder component when facing new feature distributions. Based on this, a fine-tuning scheme is proposed to make its decisions more robust and adaptable to the characteristics of new features. This avoids the catastrophic forgetting that may result from directly training the policy decoder component with limited data, achieving precise and conservative adjustments to the decision logic. Based on the decision correlation pattern, the direction in which the policy decoder component of the path planning agent should adaptively adjust is clarified.

[0092] S7.3. Following the guiding direction of the policy decoder component of the path planning agent, adjust the weight mapping relationship of the policy decoder component to the intermediate feature representation, and fine-tune the path planning agent by combining the optimized feature encoder component parameters with the adjusted policy decoder component weight mapping relationship.

[0093] Furthermore, based on the guiding direction, the weight mapping relationship of the policy decoder component is specifically adjusted. The adjustment is gentle and targeted. For example, a small weight update vector is obtained along the guiding direction and added to the current parameters of the policy decoder component; or, only the weights of the last few layers of the policy decoder component network can be adjusted. These layers are directly responsible for converting high-level features into specific action parameters. The purpose of the adjustment is to enable the policy decoder component to better understand and utilize the new feature representations generated by the optimized feature encoder component, and to bridge the decision bias that may be caused by changes in feature distribution. Combining the optimized feature encoder component parameters with the adjusted policy decoder component weight mapping relationship constitutes the complete path planning agent after fine-tuning.

[0094] Specifically, it was recognized that rapid adaptation to a new environment requires not only adjustments to the perception component but also corresponding, controlled, and coordinated changes to the decision-making component. By first exploring multiple sets of feature update paths and analyzing their decision-making correlations, and then guiding the adjustment of the policy decoder component accordingly, the entire adaptation process was ensured to be data-efficient and goal-oriented, while maintaining the stability of the original core policy logic to the greatest extent. Coordinated fine-tuning enabled the path planning agent as a whole to adapt more harmoniously to the characteristics of the new farmland, outputting more reliable and consistent operational instructions. This completed the guided adjustment of the policy decoder component and, combined with the optimized feature encoder component, achieved complete fine-tuning of the path planning agent.

[0095] S8. Using the fine-tuned path planning agent, generate and execute paths and pesticide application instructions that are both adaptable to the current farmland and optimized for future operations.

[0096] S8.1 Establish a connection between the fine-tuned path planning agent and the real-time sensor data stream of the agricultural drone. The fine-tuned path planning agent processes the current environmental observations from the sensor data stream in real time.

[0097] Furthermore, the rapidly adaptive and optimized intelligent decision-making core is deployed into a key online real-time interface. The finely tuned path planning agent, acting as a software process or neural network inference engine, is deployed on the onboard computer of the agricultural drone. It establishes a stable data connection with the drone's sensor data bus or message queue via a software interface. Real-time data streams from various sensors, including visible light cameras, multispectral sensors, lidar, inertial navigation units, and positioning modules, are pushed to the connection channel according to predefined formats and frequencies. The finely tuned path planning agent continuously monitors this data stream, receiving and analyzing current environmental observation data in real time.

[0098] Specifically, for example, camera images are decoded into pixel matrices, positioning data is parsed into latitude and longitude coordinates, and all heterogeneous data are synchronized, fused, and organized into a structured input format required by the fine-tuned path planning agent. The feature encoder component in the fine-tuned path planning agent then forward-propagates the current environmental observation data to extract time-invariant feature representations that characterize the current farmland state and UAV pose, ensuring that the agent can obtain the latest environmental information as a basis for decision-making. This realizes real-time information injection from the physical world to the digital decision model. The low latency and high reliability of the connection and processing process are the foundation for ensuring the effectiveness of subsequent real-time planning and execution, enabling the fine-tuned path planning agent to access and process real-time sensor data streams online.

[0099] S8.2 Based on current environmental observations, the fine-tuned path planning agent generates an optimized flight path that is adaptable to the current farmland. Combining the optimized flight path with the long-term optimization goals embedded in the path planning agent, a matching precise pesticide application instruction is generated.

[0100] Furthermore, the policy decoder component of the fine-tuned path planning agent receives time-invariant feature representations based on current real-time environmental observations extracted by the feature encoder component. This representation incorporates knowledge learned through small-sample fine-tuning about the unique attributes of the current new farmland (such as row spacing of specific crops, distribution of local obstacles, and unique textures of pests and diseases). This makes the generated action commands highly adaptable to the current farmland. Based on this, the policy decoder component generates an action sequence, which, when unfolded in the time dimension, constitutes an optimized flight path. For example, it consists of a series of ordered waypoint coordinates and the target speed to reach each waypoint. Since the fine-tuned path planning agent is trained through meta-reinforcement learning, its policy network parameters embed the long-term optimization objective learned through temporal rewards during the training phase.

[0101] Specifically, for example, considering the inhibitory effect of pesticide application on the development of pests and diseases in the next cycle, when generating flight paths and corresponding pesticide application instructions (such as nozzle switching timing and pesticide flow curves), the strategy decoder component not only considers how to efficiently cover the currently visible pest and disease areas, but also implicitly considers the potential positive impact of flight trajectories and pesticide application patterns on the future state of farmland. It generates a set of collaborative instructions that take into account both immediate operational effects and long-term agronomic benefits, so that the drone is not only performing a spraying operation, but also executing a sustainable field management strategy. Based on real-time perception, the fine-tuned path planning agent outputs a set of optimized flight paths and precise pesticide application instructions that integrate immediate environmental adaptability and long-term optimization goals.

[0102] S8.3. Send the optimized flight path and precise application instructions to the flight control unit and application unit of the agricultural drone.

[0103] Furthermore, this involves the crucial data distribution stage, which transforms advanced intelligent decisions into underlying executable commands. The optimized flight path generated by the fine-tuned path planning agent is typically represented as a list of waypoints linked by time or sequence indices. Each waypoint contains information such as three-dimensional position, speed, and orientation. Precision pesticide application commands may be associated with waypoints or may be an independent time-flow control sequence. The command data is packaged and sent separately to the flight control unit and pesticide application unit of the agricultural drone through communication protocols and interfaces defined within the onboard computer.

[0104] Specifically, for example, flight path data is sent to the flight control unit via a serial bus or network port, while the dosing command is sent to the dosing unit via a separate dedicated control bus. The transmission process must ensure the integrity, timing, and low latency of the commands to guarantee the accuracy of action coordination, establish a clear signal path from decision-making to execution, and ensure that high-level planning is accurately delivered to each execution endpoint. The command format needs to match the input protocols of the flight control unit and the dosing unit, which may involve coordinate transformation, unit conversion, or command encoding, reliably distributing the planning results generated by the fine-tuned path planning agent to the corresponding execution mechanisms.

[0105] S8.4 The flight control unit and the drug delivery unit work together to execute optimized flight paths and precise drug delivery commands.

[0106] Furthermore, the final step in realizing the planning instructions in the physical world embodies the closed loop of decision-making and execution. After receiving the optimized flight path, the flight control unit decomposes it into a series of low-level control commands, driving the UAV's power system (such as motors and ESCs) to operate, enabling the UAV to fly accurately according to the predetermined waypoints, speeds, and attitudes. Based on the received precise application instructions, the application unit controls the drug pump, solenoid valves, and nozzle actuators to precisely open, close, and adjust the drug flow as the UAV flies over the target area, achieving on-demand application. Precise spatiotemporal synchronization between the two units requires the flight control unit to provide real-time feedback of its position and speed status or align it with the timestamp of the application instructions, ensuring that the application action is strictly synchronized with the UAV's arrival at the specific operational location.

[0107] Specifically, for example, when the flight control unit determines that the drone is about to enter an area that needs to be sprayed, it sends a trigger signal to the spraying unit in advance. The spraying unit then starts spraying at the corresponding flow rate in the area according to the instruction, ensuring that the pesticide is accurately delivered to the target area. This achieves a deep integration of flight trajectory and spraying action, transforming the optimized decision of the fine-tuned path planning agent into actual, efficient, and precise plant protection operations. It completes the entire process from intelligent decision-making to physical execution, realizing the automation and intelligence of plant protection operations.

[0108] This embodiment also provides a deep learning-based agricultural drone spraying path planning system, including: a construction module, which in the offline stage constructs a meta-training task distribution of a virtual farmland scene, and performs architecture design and parameter optimization through a meta-learning framework to obtain a path planning agent; The training module is configured to extract time-invariant features from environmental observations, generate composite action instructions, and train the path planning agent using a meta-reinforcement learning algorithm. The data collection module deploys the trained path planning agent onto the agricultural drone and performs initial exploratory flights over the target new farmland to collect a small amount of interactive data. The fine-tuning module optimizes and updates the path planning agent by extracting its feature parameters and guides the decision-making correlations generated during the update process, thus fine-tuning the path planning agent. The execution module uses a fine-tuned path planning agent to generate and execute paths and pesticide application instructions that are both adaptable to the current farmland and optimized for future operations.

[0109] This embodiment also provides a computer device applicable to the deep learning-based agricultural drone spraying path planning method, including: a memory and a processor; the memory is used to store computer-executable instructions, and the processor is used to execute the computer-executable instructions to implement the deep learning-based agricultural drone spraying path planning method proposed in the above embodiment.

[0110] The computer device can be a terminal, comprising a processor, memory, communication interface, display screen, and input devices connected via a system bus. The processor provides computing and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system and computer programs. The internal memory provides an environment for the operation of the operating system and computer programs stored in the non-volatile storage media. The communication interface is used for wired or wireless communication with external terminals; wireless communication can be achieved through Wi-Fi, carrier networks, NFC (Near Field Communication), or other technologies. The display screen can be an LCD screen or an e-ink screen. The input devices can be a touch layer covering the display screen, buttons, a trackball, or a touchpad on the computer device's casing, or an external keyboard, touchpad, or mouse.

[0111] This embodiment also provides a storage medium storing a computer program. When executed by a processor, the program implements the deep learning-based pesticide application path planning method for agricultural drones as proposed in the above embodiments. The storage medium can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as Static Random Access Memory (SRAM), Electrically Erasable Programmable Read-Only Memory (EEPROM), Erasable Programmable Read Only Memory (EPROM), Programmable Red-Only Memory (PROM), Read-Only Memory (ROM), magnetic storage, flash memory, magnetic disk, or optical disk.

[0112] In summary, this invention constructs a meta-training task distribution that integrates a dynamic growth model, and trains a path planning agent with long-term optimization awareness through meta-reinforcement learning based on this time-series distribution. This enables the agent to acquire rapid adaptive meta-capabilities for new farmland. Through small-sample fine-tuning guided by decision correlation, the feature encoder and policy decoder are simultaneously optimized on a small amount of new data. This allows the agent to quickly adapt to the characteristics of new farmland while maintaining the stability and consistency of decision logic. The agent can generate and execute optimized paths and pesticide application instructions in real time that accurately adapt to the current real-time farmland conditions and create favorable agronomic conditions for future operation cycles. This solves the core problems of insufficient model generalization ability, slow adaptation process, and easy loss of general knowledge faced by existing technologies in new and unusual farmland scenarios. It achieves a synergistic improvement in efficiency, adaptability, and long-term agronomic benefits of agricultural drone path planning.

[0113] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A method for pesticide application path planning using agricultural drones based on deep learning, characterized in that: include, In the offline phase, a meta-training task distribution for a virtual farmland scenario is constructed, and the architecture is designed and parameters are optimized through a meta-learning framework to obtain a path planning agent. The architecture of the path planning agent is configured to extract time-invariant features from environmental observations, generate composite action instructions, and train the path planning agent using a meta-reinforcement learning algorithm. The trained path planning agent was deployed on an agricultural drone and used to conduct an initial exploratory flight over the target new farmland to collect a small amount of interactive data. The path planning agent is optimized and updated by extracting features from its parameters, and the decision-making correlations generated during the update process are used to guide the fine-tuning of the path planning agent. Using a finely tuned path planning agent, generate and execute paths and pesticide application instructions that are both adaptable to the current farmland and optimized for future operations.

2. The method for pesticide application path planning of agricultural drones based on deep learning as described in claim 1, characterized in that: In the offline phase, the meta-training task distribution for the virtual farmland scene is constructed, including the following steps: Based on the meta-training task distribution dataset, a dynamic growth model is generated for each virtual farmland scene to simulate the temporal evolution of the virtual farmland scene's state over multiple consecutive operation cycles. Based on geographic information data of real farmland, crop planting models and historical agricultural data, a virtual farmland scene description is generated through parametric modeling and data augmentation techniques. By integrating temporal evolution data of state with virtual farmland scene descriptions, a meta-training task distribution is formed.

3. The method for pesticide application path planning based on deep learning for agricultural drones as described in claim 2, characterized in that: The path planning agent is obtained through architecture design and parameter optimization using a meta-learning framework, including the following steps: Based on the meta-training task distribution, the network structure parameters of the path planning agent are initialized, and the inner and outer loop optimization processes of the meta-learning framework are executed using the meta-training task distribution. The network structure parameters of the path planning agent are updated through the inner and outer loop optimization process of the meta-learning framework. Based on the updated network structure parameters, the path planning agent is obtained.

4. The method for pesticide application path planning of agricultural drones based on deep learning as described in claim 3, characterized in that: The architecture of the path planning agent is configured to extract time-invariant features from environmental observations and generate composite action instructions, including the following steps: Configure the architecture of the path planning agent as a feature encoder component and a policy decoder component; The feature encoder component receives farmland images and environmental observation data of UAV flight status, and processes the environmental observation data to extract time-invariant feature representations. The policy decoder component outputs composite action instructions based on time-invariant feature representation.

5. The method for pesticide application path planning based on deep learning for agricultural drones as described in claim 4, characterized in that: The path planning agent is trained using a meta-reinforcement learning algorithm, including the following steps: Virtual farmland scenes are sampled from the meta-training task distribution. In the simulation environment of the current meta-task, the path planning agent interacts with the environment to generate interaction trajectory data. Based on the dynamic growth model and interaction trajectory data corresponding to the virtual farmland scene, the cumulative reward of time series reward is calculated; Based on the cumulative reward of interactive trajectory data and time-series rewards, the parameters of the path planning agent are updated by executing a meta-reinforcement learning algorithm. After sampling and training a preset number of virtual farmland scenarios in the meta-training task distribution, the trained path planning agent is obtained.

6. The method for pesticide application path planning based on deep learning for agricultural drones as described in claim 5, characterized in that: The trained path planning agent is deployed on an agricultural drone and used for initial exploratory flights over the target new farmland to collect a small amount of interaction data, including the following steps: The trained path planning agent is loaded into the onboard computer memory of the agricultural drone, and the agricultural drone is launched to fly to the airspace of the target new farmland. The plant protection drone was controlled to fly over the target new farmland in a preset initial exploration flight mode; During the initial exploration flight, the path planning agent generates tentative action commands, records the environmental observation data acquired by the sensors during the initial exploration flight, the tentative action commands generated by the path planning agent, and the corresponding real-time performance evaluation data, forming a small amount of interactive data.

7. The method for pesticide application path planning based on deep learning for agricultural drones as described in claim 6, characterized in that: The optimization and updating of parameters through feature extraction from the path planning agent includes the following steps: Environmental observation data from a small amount of interactive data is input into the feature encoder component of the path planning agent to obtain feature representations. These representations are then input into a fixed policy decoder component to generate predicted actions. The loss between the predicted actions and the actual recorded actions is calculated. The backpropagation algorithm is used to calculate the parameter gradient of the loss with respect to the feature encoder component of the path planning agent. Based on the parameter gradient, multiple sets of candidate parameter update vectors for the feature encoder component are generated. Multiple sets of candidate parameter update vectors are applied to the feature encoder component to generate multiple temporary updated versions of the feature encoder component. Multiple temporary updated versions of the feature encoder component are used to process environmental observation data in a small amount of interactive data to obtain multiple sets of intermediate feature representations. Multiple sets of intermediate feature representations are input into the policy decoder component of the path planning agent to generate multiple sets of candidate action instruction sequences. The decision consistency measure of the multiple sets of candidate action instruction sequences is evaluated, and the candidate parameter update vector with the highest decision consistency measure is selected to optimize and update the parameters of the feature encoder component of the path planning agent.

8. The method for pesticide application path planning of agricultural drones based on deep learning as described in claim 7, characterized in that: The decision-making correlations generated during the update process guide the fine-tuning of the path planning agent, including the following steps: The distribution differences among multiple sets of candidate action instruction sequences are analyzed to identify decision correlation patterns. The guidance direction for the policy decoder component of the path planning agent is determined based on the decision correlation pattern; Following the guiding direction of the policy decoder component of the path planning agent, the weight mapping relationship of the policy decoder component to the intermediate feature representation is adjusted. By combining the optimized feature encoder component parameters with the adjusted policy decoder component weight mapping relationship, the path planning agent is fine-tuned.

9. The method for pesticide application path planning of agricultural drones based on deep learning as described in claim 8, characterized in that: Using a fine-tuned path planning agent, generate and execute paths and pesticide application instructions that are both adaptable to the current farmland and optimized for future operations, including the following steps: The fine-tuned path planning agent is connected to the real-time sensor data stream of the agricultural drone, and the fine-tuned path planning agent processes the current environmental observations from the sensor data stream in real time. Based on current environmental observations, the fine-tuned path planning agent generates an optimized flight path that is adaptable to the current farmland. Combining the optimized flight path with the long-term optimization goals embedded in the path planning agent, a matching precise pesticide application instruction is generated. The optimized flight path and precise application instructions are sent to the flight control unit and application unit of the agricultural drone; The flight control unit and the drug delivery unit work together to execute optimized flight paths and precise drug delivery commands.

10. A deep learning-based agricultural drone spraying path planning system, based on the deep learning-based agricultural drone spraying path planning method according to any one of claims 1 to 9, characterized in that: This includes building modules, constructing a meta-training task distribution for a virtual farmland scene during the offline phase, and using a meta-learning framework for architecture design and parameter optimization to obtain a path planning agent; The training module is configured to extract time-invariant features from environmental observations, generate composite action instructions, and train the path planning agent using a meta-reinforcement learning algorithm. The data collection module deploys the trained path planning agent onto the agricultural drone and performs initial exploratory flights over the target new farmland to collect a small amount of interactive data. The fine-tuning module optimizes and updates the path planning agent by extracting its feature parameters and guides the decision-making correlations generated during the update process, thus fine-tuning the path planning agent. The execution module uses a fine-tuned path planning agent to generate and execute paths and pesticide application instructions that are both adaptable to the current farmland and optimized for future operations.

Images