Multi-technology fusion forest complex environment unmanned aerial vehicle coverage path planning method

Abstract

The invention discloses a multi-technology fusion forest complex environment unmanned aerial vehicle coverage path planning method, and relates to the field of unmanned aerial vehicle path planning, and the method comprises the steps: 1, constructing an unmanned aerial vehicle grid world model: abstracting a forest operation region into a two-dimensional grid map comprising an obstacle region, a landing charging region and a target coverage region; defining drone states including location, landing state, and battery level; setting an action space, and applying state constraint; step 2, establishing a Markov decision process model: defining a six-tuple including a state space, an action space, a state transfer function, a reward function, a discount factor and an observation space; repeated paths can be identified and decisions can be adjusted so that a rotation behavior of the unmanned aerial vehicle in a forest similar terrain can be effectively inhibited, path continuity can be enhanced, invalid path consumption can be reduced, an unexplored area strategy can be preferentially covered, repeated movement of a covered area can be avoided, coverage leakage can be reduced and coverage integrity of a forest target area can be enhanced.

Description

Technical Field

[0001] This invention relates to the field of unmanned aerial vehicle (UAV) path planning technology, specifically a multi-technology integrated method for UAV coverage path planning in complex forest environments. Background Technology

[0002] Coverage path planning, a core technology in the field of unmanned aerial vehicles (UAVs), has been widely applied in various scenarios such as agricultural plant protection, urban infrastructure inspection, disaster search and rescue, and environmental monitoring. Particularly in the field of forest ecological protection, it supports critical tasks such as forest resource surveys, monitoring of rare and endangered flora and fauna, and forest fire prevention patrols. Its core objective is to generate paths that efficiently cover target areas while meeting physical constraints such as UAV endurance limitations and obstacle avoidance requirements. Currently, this technology has gradually iterated from the traditional algorithm stage to the intelligent optimization stage.

[0003] Traditional algorithms are poorly adapted to complex environments. While deterministic methods can provide full coverage, they have poor dynamic adaptability and are prone to generating redundant paths in forest scenarios, making it difficult to cope with real-time changes. Heuristic algorithms do not consider battery constraints, making it difficult to support long-term operation. Some of them also have issues such as path duplication and uneven resource allocation in complex environments. Graph search solutions do not coordinate charging paths and coverage paths, limiting their practicality for large-scale operations and making it difficult to balance coverage efficiency and battery life requirements.

[0004] While DQN addresses the curse of high dimensionality in traditional Q-learning, it suffers from slow convergence in continuous action spaces and long-term tasks, making it difficult to adapt to dynamic forest changes. Some PPO algorithms lack sufficient historical awareness of local locations, easily leading to UAV path loops in complex terrain, reducing coverage efficiency, and struggling to balance path continuity and coverage integrity. High-dimensional terrain feature extraction is also inefficient. Forest scenes possess complex high-dimensional features, and traditional CNNs are prone to the vanishing gradient problem, failing to fully exploit map information, thus limiting path planning accuracy and impacting the quality and efficiency of forestry operations. Summary of the Invention

[0005] (a) Technical problems to be solved

[0006] In view of the above-mentioned shortcomings of the existing technology, the present invention provides a multi-technology integrated method for UAV coverage path planning in complex forest environments, which can effectively solve the problems of the existing technology.

[0007] (II) Technical Solution

[0008] To achieve the above objectives, the present invention provides the following technical solution:

[0009] This invention discloses a multi-technology integrated method for unmanned aerial vehicle (UAV) coverage path planning in complex forest environments, comprising the following steps:

[0010] Step 1: Construct a UAV mesh world model: Abstract the forest operation area into a two-dimensional grid map containing obstacle areas, landing and charging areas, and target coverage areas; define the UAV state, including position, landing state, and battery level; set the action space and apply state constraints;

[0011] Step 2: Establish a Markov decision process model: Define a six-tuple including: state space, action space, state transition function, reward function, discount factor, and observation space; construct the observation space using a dual-map architecture of global map and local map sequences;

[0012] Step 3: Design a composite reward function: Construct a composite reward mechanism that includes coverage progress rewards, movement and charging penalties, constraint violation penalties, and area exploration guidance rewards;

[0013] Step 4: Construct a network model based on near-end policy optimization: Design a network architecture that includes input preprocessing, a global feature extraction branch, a local temporal feature extraction branch, and a feature fusion enhancement module; the global branch extracts spatial features through convolutional layers, the local branch extracts temporal features through temporal convolution and LSTM, and the feature fusion enhancement enhances feature representation through residual connections;

[0014] Step 5, Extended State Adversarial Markov Decision Process: Define a six-tuple, where the perturbation set is defined with the true state as the center. The norm sphere perturbation set simulates signal interference and sensor noise in a forest environment, and the UAV outputs actions based on the perturbation observations.

[0015] Step 6: Implement projected gradient-directed attack: Generate adversarial observation samples through multi-step iterative optimization;

[0016] Step 7: Introduce a robust regularization term for model training: Design the total loss function, construct a robust regularization term based on KL divergence, and fuse it with the PPO standard loss to form the total loss function.

[0017] Furthermore, in step 1, the grid side length of the two-dimensional grid map is matched with the UAV's flight altitude and the camera's field of view.

[0018] The obstacle area includes: trees and ravines that cannot be flown over; the landing and charging area is a flat clearing in the forest; and the target coverage area is a set of pre-set operation grids.

[0019] The drone's status is characterized by its position, landing status, and battery level, with the battery level quantified by the remaining flight steps.

[0020] The motion space is designed for seven types of discrete motions, including east, south, west, and north movement motions, as well as takeoff, landing, and charging function motions;

[0021] State constraints include: takeoff or charging actions can only be performed during landing, movement or landing actions can only be performed during flight, and the flight state cannot enter a no-fly zone or run out of battery.

[0022] Furthermore, the state transition function of the Markov decision process model in step 2 is defined as:

[0023] Position updates follow the action-displacement mapping table. The landing state is switched only by takeoff and landing actions. The battery level update rule is to increase the amount of a single charge when charging but not exceed the maximum capacity, and to reduce the battery by 1 unit per step when not charging.

[0024] The target area is updated according to the field of view coverage rules. If the drone's position does not obstruct the grid, the grid is removed from the uncovered set.

[0025] Furthermore, in step 2, the global map architecture is a fixed-size low-resolution raster matrix that integrates static features such as obstacle distribution, no-fly zone markings, landing and charging areas, target coverage areas, and covered areas; the local map sequence includes temporal features of surrounding obstacles, uncovered target raster cells, and the historical trajectories of the last 5 steps.

[0026] Furthermore, in the reward function of step 3:

[0027] Coverage progress rewards are calculated based on the coverage increment of the target area, using the following formula:

[0028] ;

[0029] In the formula, , The number of coverage grid cells from time t to t+1, when no new area is covered. ;

[0030] Movement and charging penalties apply a negative penalty to each step of movement and invalid charging, using the following formula:

[0031] ;

[0032] In the formula, Penalty for each move. As a penalty for charging, Represents an indicator function;

[0033] The system imposes penalties for violations, setting strong negative rewards for dangerous behaviors such as entering no-fly zones or running out of battery during flight.

[0034] Furthermore, the area exploration guidance reward is designed with differentiated rewards based on the coverage status of the 8-neighbor grids of the drone's current location. If there are uncovered grids in the neighborhood and the drone can cover them after moving, the reward is 0.04; otherwise, the penalty is -0.04 or 0.

[0035] Furthermore, the network model architecture in step 4 includes:

[0036] The global feature extraction branch contains four convolutional layers, each using a 3×3 convolutional kernel, with output channels of 64, 128, 256, and 512 respectively. Each convolutional layer is followed by a ReLU activation function and a MaxPooling layer for downsampling, and finally compressed into a 512-dimensional global feature vector through global max pooling.

[0037] The local temporal feature extraction branch targets a 17×17×3 local map sequence with a length of 5. First, it applies 4 convolutional layers through the TimeDistributed layer, with the number of output channels being 32, 64, 128, and 256 respectively. Combined with ReLU activation and MaxPooling, it achieves single-frame spatial feature extraction. Then, the features of 5 time steps are input into the LSTM network to output a 64-dimensional local temporal feature vector.

[0038] The feature fusion enhancement module concatenates the global feature vector, local temporal feature vector, and 2D UAV position information into a 578-dimensional joint feature vector, which is then mapped to 512 dimensions through a fully connected layer. It is then enhanced through two layers of residual blocks. The first residual block compresses the 512-dimensional feature vector to 16 dimensions and connects it with the original feature residuals. The second residual block is processed into 32 dimensions by residual dense blocks and then connected with residuals twice. Finally, ReLU activation is used to obtain the 512-dimensional fused and enhanced feature vector.

[0039] In the output layer, the policy network maps the fused reinforcement feature vector to 7 dimensions through a fully connected layer, and then outputs the action probability distribution through softmax; the value network outputs a scalar value estimate through a fully connected layer.

[0040] Furthermore, the disturbance set in step 5 is defined as follows:

[0041] ;

[0042] In the formula, where This refers to the raw observations, i.e., the drone status information obtained from the forest environment through sensors; Indicates adversarial observation; The radius is the disturbance radius.

[0043] Furthermore, the multi-step iterative process in step 6 is as follows:

[0044] Initialize the perturbation variable to zero or a random small value, and set the number of iterations K and the step size α, where K is a positive integer, and for each iteration step k=1,2,…,K; α is the step size hyperparameter;

[0045] Calculate the loss function Regarding the disturbance variable gradient ,in V represents the negative of the expected return of the strategy, V is the value function, and the output is the long-term return estimate under the current observation.

[0046] Update the perturbation variable: , where sign( The function takes the sign of the gradient, which is used in... Efficient directional perturbation under norm constraints;

[0047] The updated perturbation variables are added back to the perturbation set, and the clip function is used to limit them. In [− , Within the scope, ensure ;

[0048] Finally, adversarial observation is generated. .

[0049] Furthermore, the expression for constructing the robust regularization term in step 7 is:

[0050] ;

[0051] In the formula, Represents the robust regularization term, D represents the data distribution of the experience replay buffer, and KL( ) represents the KL divergence. Representative in the original observation Under the given conditions, the action probability distribution output by the policy network, Representative in adversarial observation Under the given conditions, the action probability distribution output by the policy network, Represents the original observation, This represents adversarial observation.

[0052] (III) Beneficial Effects

[0053] Compared with the known prior art, the technical solution provided by this invention has the following beneficial effects:

[0054] 1. By integrating an LSTM network into the local map processing stage, high-resolution sub-map sequences containing nearly 5 historical trajectories are processed through a TimeDistributed layer to extract features. The LSTM gating mechanism is then used to remember temporal associations, which can identify repeated paths and adjust decisions. This effectively suppresses the spinning behavior of drones in similar forest terrain, improves path continuity, and reduces invalid path consumption. A residual network is introduced in the feature fusion stage, which transmits shallow gradients through skip connections to solve the gradient vanishing problem in traditional CNN processing of high-dimensional forest features. The two-layer residual block reinforcement structure also enhances the high-dimensional feature expression capability, helping the model to accurately mine key terrain information, improve the effectiveness of feature extraction, and thus accelerate model convergence, shorten the iteration cycle, and reduce training resources and time costs.

[0055] 2. By constructing a multi-dimensional composite reward function, four core items guide the drone to prioritize exploring uncovered areas: coverage progress rewards positively incentivize coverage expansion based on grid increments; movement and charging penalties reduce path redundancy and energy consumption; constraint violation penalties ensure operational safety; and area exploration guidance rewards are differentiated based on the proportion of uncovered grids in the 8-neighborhood. This mechanism enables the model to autonomously learn a strategy to prioritize covering unexplored areas, avoid repeated movement in already covered areas, reduce missed coverage, and improve the coverage integrity of forest target areas. At the same time, path redundancy is reduced, the drone's coverage range is wider with the same battery capacity, energy utilization efficiency is optimized, and it adapts to the endurance requirements of large-scale forest operations.

[0056] 3. By extending the traditional MDP to SA-MDP, define The norm sphere perturbation ensemble simulates observation bias caused by forest signal interference and sensor noise. Adversarial observations are generated using the PGD algorithm, and the perturbation range is limited by a clip operation, providing the model with realistic perturbation training samples. This allows the model to learn decision-making strategies to cope with perturbations during the training phase. The total loss function, incorporating a KL divergence robust regularization term, quantifies the difference in strategies between the original and adversarial observations, forcing decision consistency. This design effectively avoids path deviation problems caused by observation bias in forest environments, ensuring stable and reliable path planning performance even in complex perturbation scenarios. Attached Figure Description

[0057] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the accompanying drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are merely some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without any creative effort.

[0058] Figure 1 This is a schematic diagram of the process of the present invention;

[0059] Figure 2 This is a schematic diagram of the framework of the UAV mesh world model in this invention;

[0060] Figure 3 This is a diagram illustrating the architecture of the network model in this invention. Detailed Implementation

[0061] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0062] The present invention will be further described below with reference to embodiments.

[0063] This embodiment presents a multi-technology integrated UAV coverage path planning method for complex forest environments, such as... Figures 1-3 As shown, it includes the following steps:

[0064] Step 1: Construct a UAV grid world model: Abstract the forest operation area into a fixed-size two-dimensional grid map containing obstacle areas, landing and charging areas, and target coverage areas; define the UAV state, including position, landing state, and battery level; set the action space and apply state constraints to ensure flight safety; match the grid side length of the two-dimensional grid map with the UAV's flight altitude and camera field of view to ensure that a single observation fully acquires the grid cell attribute information;

[0065] The obstacle area includes: trees and ravines that cannot be flown over; the landing and charging area is a flat clearing in the forest; and the target coverage area is a set of pre-set operation grids.

[0066] The drone's status is characterized by its position, landing status, and battery level, with the battery level quantified by the remaining flight steps.

[0067] The motion space is designed for seven types of discrete motions, including east, south, west, and north movement motions, as well as takeoff, landing, and charging function motions;

[0068] State constraints include: takeoff or charging actions can only be performed during landing, movement or landing actions can only be performed during flight, and the flight state cannot enter a no-fly zone or run out of battery.

[0069] Step 2: Establish a Markov decision process model: Define a six-tuple including: state space, action space, state transition function, reward function, discount factor, and observation space; construct the observation space using a dual-map architecture of global map and local map sequence. The global map provides macroscopic environmental information, while the local map sequence provides dynamic information about the UAV's surroundings. The global map architecture is a fixed-size, low-resolution raster matrix that integrates static features such as obstacle distribution, no-fly zone markings, landing and charging areas, target coverage areas, and covered areas; the local map sequence includes temporal features of surrounding obstacles, uncovered target raster, and the historical trajectory of the last 5 steps.

[0070] The state transition function of a Markov decision process model is defined as:

[0071] Position updates follow the action-displacement mapping table. The landing state is switched only by takeoff and landing actions. The battery level update rule is to increase the amount of a single charge when charging but not exceed the maximum capacity, and to reduce the battery by 1 unit per step when not charging.

[0072] The target area is updated according to the field of view coverage rules. If the drone's position does not obstruct the grid, the grid is removed from the uncovered set.

[0073] Step 3: Design a composite reward function: Construct a composite reward mechanism that includes coverage progress rewards, movement and charging penalties, constraint violation penalties, and area exploration guidance rewards to balance coverage efficiency and flight safety; the reward function includes:

[0074] Coverage progress rewards are calculated based on the coverage increment of the target area, using the following formula:

[0075] ;

[0076] In the formula, , The number of coverage grid cells from time t to t+1, when no new area is covered. ;

[0077] Movement and charging penalties apply a negative penalty to each step of movement and invalid charging, using the following formula:

[0078] ;

[0079] In the formula, Penalty for each move. As a penalty for charging, Represents an indicator function;

[0080] The system includes strong negative rewards for violations such as entering no-fly zones or dangerous behaviors like running out of battery during flight.

[0081] The area exploration guidance reward is designed with differentiated rewards based on the coverage status of the 8 neighboring grids of the drone's current location. If there are uncovered grids in the neighborhood and the drone can cover them after moving, the reward is 0.04; otherwise, the penalty is -0.04 or 0, guiding priority to explore uncovered areas.

[0082] Step 4: Construct a network model based on near-end policy optimization: Design a network architecture that includes input preprocessing, a global feature extraction branch, a local temporal feature extraction branch, and a feature fusion enhancement module; the global branch extracts spatial features through convolutional layers, the local branch extracts temporal features through temporal convolution and LSTM, and the feature fusion enhancement enhances feature representation through residual connections;

[0083] Step 5, Extended State Adversarial Markov Decision Process: Define a six-tuple, where the perturbation set is defined with the true state as the center. The norm sphere perturbation set simulates signal interference and sensor noise in a forest environment. The UAV outputs actions based on perturbation observations, while environmental feedback is determined by the actual state. The perturbation set is defined as:

[0084] ;

[0085] In the formula, where This refers to raw observations, i.e., drone status information acquired from the forest environment via sensors, including global maps, local map sequences, location, and battery level; This indicates adversarial observation, i.e., observation after adding perturbations; The disturbance radius is a hyperparameter used to control the maximum allowable amplitude of the disturbance, simulating factors such as signal interference, sensor noise, or abrupt terrain changes in a forest environment.

[0086] Step 6: Implement projected gradient-directed attack: Generate adversarial observation samples through multi-step iterative optimization to enhance the diversity of model training data;

[0087] Step 7: Introduce a robust regularization term for model training: Design the total loss function, expressed as: In the formula, Represents the total loss function. Represents PPO cutoff loss. Represents policy entropy. The mean squared error loss represents the value network. The representative term is the KL divergence robust regularization term. , and The hyperparameters represent the parameters used to balance performance and robustness. A robust regularization term is constructed based on KL divergence and fused with the PPO standard loss to form the total loss function, optimizing network parameters and improving the model's anti-interference ability and coverage efficiency. The expression for constructing the robust regularization term is:

[0088] ;

[0089] In the formula, Represents the robust regularization term, D represents the data distribution of the experience replay buffer, and KL( ) represents the KL divergence. Representative in the original observation Under the given conditions, the action probability distribution output by the policy network, Representative in adversarial observation Under the given conditions, the action probability distribution output by the policy network, Represents the original observation, This represents adversarial observation.

[0090] Compared with existing technologies, this method employs a dual-map observation mechanism (global and local) to balance macro-planning and micro-obstacle avoidance; it designs a multi-dimensional composite reward function to effectively balance coverage efficiency and security constraints; it introduces a state adversarial training mechanism to greatly improve decision-making robustness in unfamiliar terrains and interference environments; and it integrates LSTM and residual connections in the network architecture to enhance temporal memory and feature representation capabilities. In complex forest environments, it achieves a unified approach of blind-zone-free full coverage, dynamic obstacle avoidance, and energy coordination, addressing the pain points of poor adaptability, low efficiency, and insufficient stability of traditional methods.

[0091] At other levels, in this embodiment, such as Figure 3 As shown, the network model architecture includes:

[0092] The global feature extraction branch contains four convolutional layers, each using a 3×3 convolutional kernel, with output channels of 64, 128, 256, and 512 respectively. Each convolutional layer is followed by a ReLU activation function and a MaxPooling layer for downsampling, and finally compressed into a 512-dimensional global feature vector through global max pooling.

[0093] The local temporal feature extraction branch targets a 17×17×3 local map sequence with a length of 5. First, it applies 4 convolutional layers through the TimeDistributed layer, with the number of output channels being 32, 64, 128, and 256 respectively. Combined with ReLU activation and MaxPooling, it achieves single-frame spatial feature extraction. Then, the features of 5 time steps are input into the LSTM network to output a 64-dimensional local temporal feature vector.

[0094] The feature fusion enhancement module concatenates the global feature vector, local temporal feature vector, and 2D UAV position information into a 578-dimensional joint feature vector, which is then mapped to 512 dimensions through a fully connected layer. It is then enhanced through two layers of residual blocks. The first residual block compresses the 512-dimensional feature vector to 16 dimensions and connects it with the original feature residuals. The second residual block is processed into 32 dimensions by residual dense blocks and then connected with residuals twice. Finally, ReLU activation is used to obtain the 512-dimensional fused and enhanced feature vector.

[0095] In the output layer, the policy network maps the fused reinforcement feature vector to 7 dimensions through a fully connected layer, and then outputs the action probability distribution through softmax; the value network outputs a scalar value estimate through a fully connected layer.

[0096] This embodiment provides a multi-step iterative process as follows:

[0097] Initialize the perturbation variable to zero or a random small value, and set the number of iterations K and the step size α, where K is a positive integer. For each iteration step, k = 1, 2, ..., K, usually taken as 5 to 10 to balance computational efficiency and adversarial sample quality; α is the step size hyperparameter, which controls the magnitude of each gradient update.

[0098] Calculate the loss function Regarding the disturbance variable gradient ,in V represents the negative of the expected policy return, and V is the value function that outputs the long-term return estimate under the current observations. This loss function aims to minimize the expected policy return, thereby generating adversarial examples that maximize policy bias.

[0099] Update the perturbation variable: , where sign( The function takes the sign of the gradient, which is used in... Efficient directional perturbation under norm constraints;

[0100] The updated perturbation variables are added back to the perturbation set, and the clip function is used to limit them. In [− , Within the scope, ensure ;

[0101] Finally, adversarial observation is generated. This adversarial observation is used in state adversarial training. By comparing the policy distribution under the original observation and the adversarial observation, the KL divergence robust regularization term is calculated to enhance the model's robustness to uncertainties in complex forest environments.

[0102] In summary, this invention constructs a grid-world model, abstracting the forest environment into a structured representation that includes obstacles, charging zones, and target coverage areas, enabling effective planning without the need for high-precision maps. Furthermore, by combining a global-local dual-map observation architecture, it ensures both the rationality of the macroscopic path and the accuracy of local obstacle avoidance. Employing a PPO-based deep reinforcement learning framework, it guides the UAV's intelligent decision-making through a multi-dimensional composite reward function, effectively balancing coverage integrity, path efficiency, and safety constraints.

[0103] A state adversarial training mechanism is introduced, and adversarial examples are generated through the SA-MDP framework and PGD attack. Combined with KL divergence regularization, the model has strong resistance to environmental disturbances such as sensor noise and signal interference. The LSTM module is integrated into the network architecture, which can effectively capture the temporal correlation information of local map sequences.

[0104] Through the synergistic effect of residual networks and adversarial training, the trained model can quickly adapt to new, unseen forest terrains, significantly reducing the need for retraining. It is particularly suitable for large-scale forest resource monitoring applications. By using a reward function, it simultaneously optimizes coverage integrity, path length, energy consumption, and safety performance, achieving a balance between multiple objectives and overcoming the performance bottleneck caused by single-objective optimization.

[0105] 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 the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions will not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A multi-technology fusion forest complex environment unmanned aerial vehicle coverage path planning method, characterized in that, The method comprises the following steps: Step 1, constructing a UAV grid world model: abstracting the forest operation area into a two-dimensional grid map containing obstacle areas, landing charging areas and target coverage areas; defining the UAV state including position, landing state and battery level; Set the action space and impose state constraints; Step 2, establishing a Markov decision process model: defining a six-tuple including: state space, action space, state transition function, reward function, discount factor and observation space; using a dual map architecture of global map and local map sequence to construct the observation space; Step 3, designing a composite reward function: constructing a composite reward mechanism containing coverage progress reward, movement and charging penalty, constraint violation penalty and region exploration guide reward; Step 4, constructing a network model based on proximal policy optimization: designing a network architecture containing input preprocessing, global feature extraction branch, local time series feature extraction branch and feature fusion reinforcement module; the global branch extracts spatial features through convolutional layers, the local branch extracts time series features through time series convolution and LSTM, and the feature fusion reinforcement strengthens the feature expression through residual connection; Step 5, Extended State-Antagonistic Markov Decision Process: Define a sextuple where the disturbance set is defined to be centered around the true state a norm-ball disturbance set that models signal interference and sensor noise in a forest environment, and the UAV outputs actions based on disturbance observations; Step 6, implementing a projected gradient directed attack: generating an adversarial observation sample through multi-step iterative optimization; Step 7, introducing a robust regularization term for model training: designing a total loss function; constructing a robust regularization term based on KL divergence, and combining it with the standard PPO loss to form the total loss function. 2.The multi-technology integrated forest complex environment UAV coverage path planning method according to claim 1, characterized in that, The grid length of the two-dimensional grid map in step 1 matches the flight height of the UAV and the camera field of view angle; The obstacle area includes: trees and ravines that cannot be flown over, the landing charging area is a flat open space in the forest, and the target coverage area is a set of preset operation grids; The UAV state is represented by position, landing state and battery level, and the battery level is quantified by the number of remaining flight steps; The action space is designed as seven types of discrete actions, including east, south, west, north movement actions, and take-off, landing and charging function actions; The state constraints include: only landing can execute take-off or charging actions, only flying can execute movement or landing actions, and the flying state cannot enter the no-fly zone or run out of battery.

3. The multi-technology integrated forest complex environment UAV coverage path planning method according to claim 1, characterized in that, The state transition function of the Markov decision process model in step 2 is defined as: The position update follows the action and displacement mapping table, the landing state is switched only by take-off and landing actions, and the battery level update rule is to increase the single charging amount when charging but not more than the maximum capacity, and to reduce 1 unit of battery per step when moving. The target area update is based on the field of view coverage rule, if the UAV position has no obstacle blocking the grid, the grid is removed from the uncovered set.

4. The multi-technology integrated forest complex environment UAV coverage path planning method according to claim 1, characterized in that, The global map architecture in step 2 is a fixed-size low-resolution grid matrix, integrating static features such as obstacle distribution, no-fly zone markers, landing charging areas, target coverage areas and covered areas; the local map sequence contains the surrounding obstacles, uncovered target grids and the time series features of the last 5 steps of the trajectory.

5. The multi-technology integrated forest complex environment UAV coverage path planning method according to claim 1, characterized in that, In the reward function in step 3: The coverage progress reward is calculated based on the target area coverage increment to provide positive incentive, and the formula is: ； In the formula, , is the number of covered grids from time t to t+1, when no new area is covered ; The movement and charging penalty sets negative penalties for each step of movement and invalid charging, and the formula is: ； wherein is a penalty for each move, is a penalty for charging, represents an indicator function; The constraint violation penalty sets a strong negative reward mechanism for entering a forbidden area or a dangerous behavior of battery depletion in flight.

6. The multi-technology integrated forest complex environment UAV coverage path planning method according to claim 1, characterized in that, The region exploration guidance reward is designed based on the coverage state of the 8-neighbor grid of the current position of the UAV, and if there is an uncovered grid in the neighborhood that can be covered after the UAV moves, a reward of 0.04 is given, otherwise a penalty of -0.04 or 0 is given.

7. The multi-technology integrated forest complex environment UAV coverage path planning method according to claim 1, characterized in that, The content of the network model architecture in step 4 includes: The global feature extraction branch contains 4 convolutional layers, each using a 3×3 convolutional kernel, with the number of output channels being 64, 128, 256, and 512 in turn, and after each convolution, a ReLU activation function and a MaxPooling layer are concatenated for downsampling, and finally compressed into a 512-dimensional global feature vector through global max pooling; The local temporal feature extraction branch is for a 17×17×3, length 5 local map sequence, which first applies 4 layers of convolution through the TimeDistributed layer, with the number of output channels being 32, 64, 128, and 256 in turn, and cooperates with ReLU activation and MaxPooling to realize single-frame spatial feature extraction, and then inputs the 5 time step features into the LSTM network to output a 64-dimensional local temporal feature vector; The feature fusion enhancement module concatenates the global feature vector, the local temporal feature vector, and the 2-dimensional UAV position information into a 578-dimensional joint feature vector, and then maps it to 512 dimensions through a fully connected layer, and then strengthens it through two residual blocks, the first residual block compresses 512 dimensions to 16 dimensions and connects with the original feature residual, the second residual block is processed by a residual dense block to 32 dimensions and then connected with the residual again, and finally a 512-dimensional fusion enhanced feature vector is obtained through ReLU activation; In the output layer, the policy network maps the fusion enhanced feature vector to 7 dimensions through a fully connected layer, and then outputs the action probability distribution through softmax; the value network outputs a scalar value estimate through a fully connected layer.

8. The multi-technology integrated forest complex environment UAV coverage path planning method according to claim 1, characterized in that, The perturbation set in step 5 is defined as: ； wherein represents the original observation, i.e. the drone state information acquired from the forest environment by the sensors; represents the adversarial observation; is the perturbation radius.

9. The multi-technology integrated forest complex environment UAV coverage path planning method according to claim 1, characterized in that, The multi-step iteration process in step 6 is: Initialize the perturbation variable to zero or a small random value, and set the iteration number K and the step size α, where K is a positive integer, and for each iteration step k=1,2…,K; α is the step size hyperparameter; Computing the loss function With respect to the perturbed variable Gradient of Where Represents the negative of the policy expected return, V is the value function, and outputs an estimate of the long-term return given the current observation; Updating the perturbation variable: where the sign( ) function takes the sign of the gradient, used for efficient directional perturbation under norm constraints; The updated perturbation variable is thrown back into the set of perturbations, limited by the clip function In the range [− , ] it is ensured that ; Final generated adversarial observation .

10. The multi-technology integrated forest complex environment UAV coverage path planning method according to claim 1, characterized in that, The construction expression of the robust regularization term in step 7 is: ； wherein represents a robust regularizer, D represents a data distribution of the experience replay buffer, KL( ) represents a KL divergence, represents an action probability distribution output by the policy network under the condition of the original observation represents an action probability distribution output by the policy network under the condition of the adversarial observation represents an action probability distribution output by the policy network under the condition of the original observation represents an action probability distribution output by the policy network under the condition of the adversarial observation represents the original observation, represents the adversarial observation.

Images