A method for planning an active inspection path of a unmanned aerial vehicle for saline paddy fields

Abstract

The application relates to the fields of intelligent agriculture, low-altitude agricultural inspection, unmanned aerial vehicle path planning, remote sensing monitoring and artificial intelligence, and discloses a kind of unmanned aerial vehicle active inspection path planning method for saline paddy field, including determining target saline paddy field area, and discretization is grid environment, obtains multi-source data in different grid environment;According to the multi-source data, the water salt stress related characteristic index is extracted, the water salt stress risk value and the uncertainty value of different grids are calculated, the water salt stress risk value and the uncertainty value are mapped to the grid environment, the water salt stress risk map and the uncertainty map are obtained;The unmanned aerial vehicle plans the inspection path according to the water salt stress risk map and the uncertainty map, the application provides a kind of unmanned aerial vehicle active inspection path planning method for saline paddy field, can reduce the repeated inspection of low-risk area, reduce invalid voyage.

Description

Technical Field

[0001] This invention relates to the fields of smart agriculture, low-altitude agricultural inspection, drone path planning, remote sensing monitoring and artificial intelligence, and in particular to a drone active inspection path planning method for saline-alkali rice fields. Background Technology

[0002] Saline-alkali paddy fields are easily affected by factors such as soil salinity, abnormal moisture, topographic waterlogging, and uneven crop growth during the production process, resulting in areas with varying degrees of water and salt stress. Regions such as Zhenlai County and Da'an City in western Jilin Province have typical saline-alkali paddy field application scenarios, making them suitable for low-altitude intelligent agricultural inspection and precision field management.

[0003] Currently, drones are mainly used in agriculture for plant protection spraying, crop condition monitoring, remote sensing mapping, and fixed-route inspections. Existing methods typically employ regular routes, full-coverage routes, or manually preset waypoints. While these methods can cover an entire field, they struggle to dynamically adjust inspection priorities based on differences in water and salt stress risks within the farmland.

[0004] In the scenario of monitoring water and salt stress in saline-alkali paddy fields, high-risk areas and areas where remote sensing assessment is unstable typically require more low-altitude verification; while repeated inspections of low-risk areas, areas already inspected, or areas with low information gain would result in wasted flight range and power. Therefore, existing regular or full-coverage flight routes, under conditions of limited power and flight range, are prone to insufficient inspection of high-risk and highly uncertain areas, as well as repeated inspections of low-value areas. Summary of the Invention

[0005] To address the aforementioned technical problems, this invention provides a method for proactive inspection path planning of saline-alkali paddy fields using unmanned aerial vehicles (UAVs), which can reduce repetitive inspections in low-risk areas and reduce ineffective flight paths.

[0006] This invention provides a method for planning active inspection paths using unmanned aerial vehicles (UAVs) in saline-alkali paddy fields. The method includes: determining the target saline-alkali paddy field area; discretizing the target saline-alkali paddy field area into a grid environment and acquiring multi-source data in different grid environments; extracting water and salt stress-related feature indicators from the multi-source data and calculating the water and salt stress risk value and uncertainty value for different grids; mapping the water and salt stress risk value and uncertainty value to the corresponding grid environment to obtain a water and salt stress risk map and an uncertainty map; planning an inspection path using the water and salt stress risk map and uncertainty map; recommending key verification sampling points based on the coverage of the inspection path and the water and salt stress risk map; and updating the water and salt stress risk map and uncertainty map based on UAV inspection feedback data and ground measurement data.

[0007] Optionally, when the UAV plans its inspection path based on the risk map and uncertainty map, the inspection priority of a candidate grid cell that meets the safe return-to-home condition is determined by a weighted sum of the grid cell's risk value, uncertainty value, and degree of insufficient inspection coverage. The calculation formula is as follows: , in, Indicates the first Inspection priority of each grid cell; Indicates the risk value of water and salt stress; Indicates the value of uncertainty; Indicates the first The degree of inspection coverage of a grid cell or its neighborhood; Indicates the degree of inadequacy in inspection coverage; , and These are the weighting coefficients, and Risk value weight The primary weight is determined by risk value; drones prioritize access to grid cells with high inspection priority by using risk value as the primary factor and uncertainty value and the degree of insufficient inspection coverage as secondary factors.

[0008] Optionally, the safe return-to-home condition is: the drone's current remaining battery power is greater than the sum of the battery power required to return to the return point after traveling from the current location to the candidate grid cell and the safe redundancy battery power; candidate grid cells that do not meet the safe return-to-home condition are not included in the calculation of inspection priority.

[0009] Optionally, the inspection priority is determined through a reinforcement learning model for path planning. The reward function of the reinforcement learning model is: , in, express Momentary reward value; Indicates information gain reward; Indicates a reward for coverage of areas with high water and salt stress risk; Indicates a reward for accessing the new region; This indicates a reward for highlighting the value of key sampling points. Indicates path length penalty; Indicates energy consumption penalty; Indicates a penalty for repeated access; This indicates safety penalties such as crossing the boundary, entering a no-fly zone, insufficient battery power, or inability to return to base. , , , , , , and The weighting coefficients are used in the reward function. Information gain rewards are determined based on the uncertainty of a previously unvisited grid cell; higher uncertainty results in a larger reward. High-risk area coverage rewards provide a positive reward when the risk value of a grid cell accessed by the drone exceeds a preset high-risk threshold; higher risk values ​​result in a larger reward. New area access rewards provide a fixed positive reward each time the drone first accesses a grid cell. Path length penalties and energy consumption penalties are determined based on the distance the drone travels per step and its energy consumption. Repeated access penalties are determined based on the historical number of times the current grid cell has been accessed; more accesses result in a larger penalty. Safety penalties are triggered when the drone crosses boundaries, enters a no-fly zone, or has insufficient remaining battery power for a safe return. The drone selects an action to perform based on its current state using a reinforcement learning model. After performing the action, it receives a reward and moves to the next state. The path planning model is trained by maximizing the accumulated expected reward.

[0010] Optionally, the formula for calculating the risk value of water and salt stress is as follows: , in, Indicates the first Water and salt stress risk value for each grid cell; Indicates the salt risk factor; Indicates vegetation stress risk factors; Indicates the risk factors for abnormal moisture levels; Indicates the risk factor for heat stress; Indicates a priori risk factors for saline-alkali land or terrain; , , , and Let represent the weight coefficients of each factor, and .

[0011] This optimal solution integrates multi-dimensional characteristics such as salinity, vegetation, water, heat, and topography, overcoming the one-sidedness of evaluating the stress status of saline-alkali paddy fields with a single indicator, and improving the comprehensiveness and accuracy of risk identification.

[0012] Optionally, the formula for calculating the uncertainty value is: , in, For the first The uncertainty value of each grid cell. Indicates standard deviation, Normalization is indicated; the uncertainty value represents the degree of consistency of the judgment results of each risk factor for the same grid cell. The larger the value, the higher the dispersion between each risk factor and the more unstable the remote sensing judgment.

[0013] Optionally, the drone will trigger automatic return when its remaining battery power is less than or equal to the sum of the battery power required to return to the home point from the current location and the safety margin battery power. , in, This is the current remaining battery level; For the drone from its current location The amount of electricity required to return to the starting point; For safety, redundant power supply.

[0014] Optionally, recommended key verification sampling points include: the verification priority of candidate sampling points is calculated by weighted summation after assigning positive weights to the water and salt stress risk value, uncertainty value, and degree of insufficient inspection coverage of the candidate point; the spatial distance between the candidate sampling point and existing sampling points, selected verification points, or inspected paths is used as the minimum spatial interval constraint for the subsequent screening stage; all candidate sampling points are sorted from high to low verification priority, and then screened sequentially according to the minimum spatial interval constraint to output a number of spatially dispersed key verification sampling points.

[0015] Optionally, the review priority of candidate sampling points is calculated based on a comprehensive assessment of the candidate sampling point's water and salt stress risk value, uncertainty value, degree of insufficient inspection coverage, spatial dispersion distance between the candidate sampling point and existing sampling points or selected review points, and reachability distance between the candidate sampling point and the active inspection path. The calculation formula is as follows: , in, Indicates the first The review priority of each candidate sampling point; This indicates the water and salt stress risk value at that point; This represents the uncertainty value at that point; This indicates the level of inspection coverage in the area where the point is located; Indicates the degree of inadequacy in inspection coverage; This represents the normalized spatial dispersion distance between this point and existing sampling points or selected verification points; This represents the normalized reachability distance between the point and the active inspection path or the nearest reachable waypoint; , , , and These are the weighting coefficients.

[0016] Optionally, updating the water and salt stress risk map based on UAV inspection feedback data and ground-based measured data includes: when converting UAV inspection feedback data into new risk observation values, performing orthorectification, geometric correction, and radiometric correction on the UAV imagery, and extracting low-altitude vegetation index, moisture index, salinity-related index, or texture features at a spatial scale consistent with the grid environment, and obtaining UAV-observed risk values ​​according to the same risk factor conversion rules as the initial risk map; normalizing the conductivity, pH, soil moisture content, or water conductivity in the ground verification data to obtain ground-observed risk values; when both UAV-observed and ground-observed risk values ​​exist simultaneously, obtaining new risk observation values ​​by weighting them according to their respective confidence levels, and updating the water and salt stress risk map using a confidence-weighted update method. , in, This indicates the updated risk value for water and salt stress. This indicates the water and salt stress risk value before the update; This indicates a new risk observation obtained from drone inspection data, ground verification data, or a confidence-weighted average of both. This indicates an update to the weights; Indicates the confidence level of the new observation data; This indicates the confidence level of the original risk map.

[0017] Compared with existing technologies, the technical solution provided by this invention has the following advantages: It can improve the inspection efficiency of high-risk areas. This invention plans the UAV inspection path based on the water and salt stress risk map, enabling the UAV to prioritize access to high-risk areas and avoiding the problem of traditional full-coverage routes inspecting all areas on an average basis. Through new area access rewards, path length penalties, and repeated access penalties, this invention can reduce repeated inspections of low-risk areas and reduce ineffective flight paths. This invention incorporates remaining battery power, maximum range, and safe return constraints into the reinforcement learning decision-making process, enabling the UAV to prioritize monitoring of key areas with limited battery power. By introducing information gain rewards and a high-uncertainty area priority inspection mechanism, this invention enables the UAV to acquire more valuable agricultural information. This invention not only outputs the UAV inspection path but also recommends key verification sampling points based on risk and uncertainty values, providing a basis for subsequent ground sampling of EC, pH, soil moisture content, etc. This invention can update the water and salt stress risk map based on UAV inspection results and ground measurement results, realizing a closed loop from risk identification, active inspection, verification sampling to risk updating. Attached Figure Description

[0018] Figure 1 The overall flowchart of a method for active inspection path planning of saline-alkali paddy fields by unmanned aerial vehicles (UAVs) provided in an embodiment of the present invention is shown below. Figure 2 A flowchart for constructing a water and salt stress risk map is provided in this embodiment of the invention. Figure 3 This is a schematic diagram of a grid environment for active inspection by a drone, provided as an embodiment of the present invention. Figure 4 A flowchart of a reinforcement learning-based active inspection path planning method is provided in this embodiment of the invention. Figure 5 This is a comparison diagram of an active inspection path and a traditional full-coverage path provided in an embodiment of the present invention. Detailed Implementation

[0019] The following detailed description of a specific embodiment of the present invention is provided in conjunction with the accompanying drawings. However, it should be understood that the scope of protection of the present invention is not limited to the specific embodiment.

[0020] In the description of this invention, it should be understood that the terms "center," "longitudinal," "lateral," "length," "width," "thickness," "upper," "lower," "front," "rear," "left," "right," "vertical," "horizontal," "top," "bottom," "inner," "outer," "axial," "radial," and "circumferential" indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. They are only for the convenience of describing the technical solution of this invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on this invention.

[0021] The present invention will be described below through several specific embodiments. To keep the following description of the embodiments clear and concise, detailed descriptions of known functions and components may be omitted. When any component of an embodiment of the present invention appears in more than one drawing, the component may be represented by the same reference numerals in each drawing.

[0022] Figure 1 This is a flowchart illustrating the overall process of a drone-based active inspection path planning method for saline-alkali paddy fields, as provided in an embodiment of the present invention. Figure 2 This invention provides a flowchart for constructing a water and salt stress risk map. Figure 3 This is a schematic diagram of a grid environment for active inspection by a drone, provided in an embodiment of the present invention. Figure 4 This invention provides a flowchart of a reinforcement learning-based active inspection path planning method. Figure 5 This is a comparison diagram of an active inspection path and a traditional full-coverage path provided in an embodiment of the present invention.

[0023] like Figure 1 As shown in the figure, this embodiment provides a method for proactive drone inspection path planning in saline-alkali paddy fields. This method constructs a dual perception map of risk and uncertainty, driving the drone to perform adaptive inspections and form a data closed loop. Specifically, it includes the following steps.

[0024] Identify the target saline-alkali paddy field area. Specifically, this step aims to clarify the spatial boundaries and object attributes of the inspection operation. The target saline-alkali paddy field area typically refers to farmland plots with typical salinization characteristics and planted with rice crops. In practical applications, this area can be delineated based on administrative divisions, geographical coordinates, or specific agricultural management units. For example, contiguous paddy fields with high soil salinization and significant differences in rice growth can be selected as the target area. Determining the target area not only limits the geographical scope of subsequent data collection and processing but also provides physical constraints for path planning, ensuring that the UAV performs its mission within legal and safe airspace.

[0025] The target saline-alkali paddy field area is discretized into a grid environment to obtain multi-source data in different grid environments.

[0026] Discretization into a grid environment refers to dividing a continuous farmland geographic space into several regular geometric units, such as square or hexagonal grids, according to a preset spatial resolution. This discretization process is not a simple geometric division, but rather aims to achieve a balance between preserving spatial details and computational efficiency. Each grid cell, as the smallest granularity for data fusion and decision execution, carries all environmental attribute information for that location. The choice of grid size should comprehensively consider the effective swath width of the UAV sensor, its flight altitude, and the average scale of saline-alkali patches; for example, it can be set to 10m × 10m to 50m × 50m. When acquiring multi-source data, data from various channels, such as satellite remote sensing, meteorological monitoring stations, digital elevation models, and historical soil survey data, are aggregated for each grid cell. This multi-source data constitutes the basic information layer for subsequent risk assessment, enabling each grid to possess multi-dimensional environmental description capabilities.

[0027] Water and salt stress-related characteristic indicators were extracted from multi-source data, and water and salt stress risk values ​​and uncertainty values ​​for different grids were calculated. The water and salt stress risk values ​​and uncertainty values ​​were then mapped to the corresponding grid environments to obtain water and salt stress risk maps and uncertainty maps.

[0028] Specifically, water and salt stress-related characteristic indicators include, but are not limited to, vegetation indices, moisture indices, salinity indices, surface temperature, and topographic factors. These indicators reflect the growth status and environmental pressure of paddy fields from different perspectives. Based on this, this embodiment constructs a dual-representation map. The water and salt stress risk map is used to quantify the severity of water and salt stress problems in each grid; the higher the value, the greater the likelihood of stress or the more severe the damage in that area. The uncertainty map, on the other hand, is used to quantify the credibility of the current risk assessment, reflecting the consistency or completeness of information among multi-source data. For example, when the vegetation index of a grid shows normal but the salinity index shows abnormal, although the overall risk value may be at a moderate level, its uncertainty value will increase significantly, indicating that there is a contradiction in the remote sensing interpretation results for that area, requiring further verification. This dual-representation mechanism enables the system not only to identify "where there is a problem" but also "where it is unclear," providing a scientific basis for subsequent differentiated inspection strategies.

[0029] The drone plans the inspection route based on the water and salt stress risk map and uncertainty map; based on the coverage of the inspection route and the water and salt stress risk map, it recommends key sampling points for verification.

[0030] This step highlights the essential difference between proactive inspection and traditional pre-defined route inspection. Traditional inspections often employ full coverage or fixed route patterns, treating all areas equally; however, the inspection path planning in this embodiment is a data-driven, dynamic decision-making process. The drone prioritizes flying to grid areas with high risk or uncertainty, while lowering the priority of low-risk and confirmed safe areas or skipping them entirely, thereby maximizing information acquisition efficiency within a limited flight time. Simultaneously, based on the actual coverage and risk distribution of the inspection path, the system automatically selects the most valuable points for verification as key sampling points. These sampling points are typically areas with extremely high risk but lacking ground-based verification, or areas with conflicting remote sensing signals requiring urgent on-site verification. Recommending sampling points bridges the gap between aerial remote sensing and ground operations, enabling inspection results to directly guide agronomists' field sampling work.

[0031] The water and salt stress risk map and uncertainty map are updated based on the feedback data from drone inspections and ground measurement data.

[0032] Specifically, this is a crucial link in a feedback loop. Low-altitude, high-resolution imagery acquired by drones during inspections, along with ground-based measurement data (such as soil conductivity and pH) at recommended sampling points, are considered more accurate true observations than the initial multi-source data. These newly acquired high-confidence data are used to revise and iterate the existing water and salt stress risk and uncertainty maps. For example, if ground measurements confirm that a high-uncertainty area is actually low-risk, the risk and uncertainty values ​​for that area are reduced in the updated map; conversely, if new high-risk patches are discovered, the risk rating of the corresponding grid is increased. Through this continuous dynamic updating, the risk map can increasingly approximate the actual field conditions as the number of inspections increases, while also providing more accurate prior knowledge for the planning of the next round of inspection routes, thereby achieving adaptive evolution and continuous performance improvement of the monitoring system.

[0033] In this embodiment, the specific calculation method for the risk value and uncertainty value of water and salt stress is explained in detail. For example... Figure 2 As shown, this calculation process, following the feature index extraction, is the core step in constructing the water and salt stress risk map. Specifically, the formula for calculating the water and salt stress risk value is:

[0034] , in, Indicates the first Water and salt stress risk value for each grid cell; Indicates the salt risk factor; Indicates vegetation stress risk factors; Indicates the risk factors for abnormal moisture levels; Indicates the risk factor for heat stress; Indicates a priori risk factors for saline-alkali land or terrain; , , , and Let each represent the weight coefficient of the factor, and satisfy the following conditions: .

[0035] The above formula quantifies the complex saline-alkali paddy field environment into a single risk value through a multi-dimensional factor weighted fusion approach. The selection of each risk factor has clear agronomic and physical basis, collectively constituting a comprehensive characterization of water and salt stress. Specifically, the salinity risk factor... It is an indicator that directly reflects the degree of soil salinization, and can be obtained through soil electrical conductivity inversion data or salinity index. The higher the value, the greater the threat of salt damage. Vegetation stress risk factors This reflects the physiological response of crops to stress. Since suppressed growth in rice is usually the most direct and comprehensive manifestation of water and salt stress, this factor plays an important role in the model. In actual treatments, if a higher original vegetation index (such as NDVI) value indicates better growth, reverse normalization is required to convert it into a risk value; that is, the worse the vegetation, the higher the risk. Water anomaly risk factor. This method is used to capture the unique water imbalance conditions in saline-alkali paddy fields, including both drought and waterlogging in low-lying areas. It is typically determined by calculating the deviation of the current water index from a suitable water reference value. Heat stress risk factors. Based on surface temperature data, since water stress often leads to a decrease in crop transpiration capacity and thus an increase in canopy temperature, high temperature anomalies can serve as an auxiliary criterion for latent stress. Saline-alkali land or topographical prior risk factors. This incorporates static environmental information such as relative elevation, slope, runoff conditions, and historical salinity distribution maps to characterize background risks that are not easily affected by seasonal changes but have a long-term impact on water and salt transport, such as secondary salinization that is more likely to occur in low-lying areas.

[0036] Regarding weighting coefficients , , , and The determination of weights is not limited to a single method, but aims to support adaptive adjustments in different scenarios. One implementation method is the expert experience method, which assigns scores to the importance of each factor based on the knowledge of local agricultural experts. Another implementation method is the objective weighting method, such as the entropy weighting method or the coefficient of variation method, which automatically calculates weights based on the information content or volatility of each factor's data, avoiding subjective bias. Yet another implementation method is to utilize historical ground-based measured samples and train machine learning algorithms such as random forests and extreme gradient boosting to obtain the optimal weight combination. For example, in a typical saline-alkali rice-growing area, the weight configuration obtained after training might be... , , , , This reflects the risk characteristics of the region, which are primarily driven by salinity and vegetation response.

[0037] Furthermore, to identify areas where remote sensing assessments are unstable or require further verification by drones, this embodiment also constructs an uncertainty map. The formula for calculating the uncertainty value is:

[0038] , in, For the first The uncertainty value of each grid cell. Indicates standard deviation, Normalization is indicated; the uncertainty value represents the degree of consistency of the judgment results of each risk factor for the same grid cell. The larger the value, the higher the dispersion between each risk factor and the more unstable the remote sensing judgment.

[0039] It is important to emphasize that the "uncertainty" in this application does not refer to sensor measurement errors or data noise, but rather to the degree of semantic conflict or cognitive disagreement among multiple sources of information. In the monitoring of saline-alkali paddy fields, single indicators often have limitations or ambiguity; a judgment is only credible when multiple independent indicators point to a consistent conclusion. For example, the vegetation index of a certain grid cell shows normal growth (low risk), but the salinity index shows severe salinization (high risk), while the topographic factor indicates that the area is a salt-accumulating depression (high risk). In this case, although the weighted comprehensive risk value... It may be at a moderate level, but due to the large differences between the factors, the calculated standard deviation is... It will be significantly higher, thus leading to uncertainty values. The increase in uncertainty clearly indicates to the system that the current remote sensing interpretation results are contradictory, possibly due to misjudgments caused by the planting of salt-tolerant varieties, image pixel mixing effects, or special micro-topography. It is necessary to dispatch drones for low-altitude, high-resolution verification to eliminate ambiguity. Conversely, if all factors show low risk or all show high risk, the standard deviation is small, the uncertainty is low, indicating that the current judgment has a high degree of confidence.

[0040] In practical implementation, the formula... The function maps the standard deviation to a uniform dimensional range (e.g., between 0 and 1) to facilitate subsequent fusion calculations with the hazard value. As a preferred implementation, the Min-Max normalization method can be used. ,in and These represent the minimum and maximum standard deviations across the entire study area, respectively. Alternatively, Z-score standardization followed by a Sigmoid function can be used to accommodate cases where the data distribution contains extreme outliers. Through this mechanism, the uncertainty map and the risk map complement each other, jointly supporting the dynamic balance between "exploration" and "utilization" strategies in subsequent proactive inspection path planning.

[0041] In this embodiment, the inspection path planning decision-making mechanism based on priority evaluation and reinforcement learning is described in detail. This mechanism integrates explicit rule guidance and implicit learning optimization, aiming to solve the core technical problem of how to balance risk investigation and information exploration for UAVs with limited power. Figure 3 and Figure 4As shown, this decision-making process is not a direct application of a single algorithm, but follows a hierarchical logic of "safety pre-filtering - priority explicit evaluation - reinforcement learning dynamic optimization".

[0042] First, the safe return-to-home condition is a rigid prerequisite for path planning. Specifically, when a UAV plans its inspection path based on the risk map and uncertainty map, only candidate grid cells that meet the safe return-to-home condition are allowed to enter the subsequent priority calculation or action selection stage. This safe return-to-home condition is defined as: the UAV's current remaining battery power is greater than the sum of the battery power required to return to the return point after traveling from the current location to the candidate grid cell, and the safety redundancy battery power; candidate grid cells that do not meet the safe return-to-home condition are not included in the inspection priority calculation. Within this constraint, This indicates the drone's current remaining battery power. Indicates the drone's current location The amount of electricity required to return to the starting point. This represents the safety redundancy of battery power. It is particularly important to emphasize that this application designs the safety constraint as a "candidate grid filter" rather than a penalty term in the reinforcement learning reward function. This is because in the early stages of reinforcement learning training, the agent may attempt high-risk actions to explore the boundaries. If only negative rewards are used to suppress dangerous behavior, not only will the convergence speed be slow, but there is still a risk of battery depletion due to model generalization errors during real-world deployment. By making the safety condition a hard threshold from the outset, the system directly eliminates grid nodes that, once executed, cannot safely return when generating the candidate action space at each time step. Figure 3 As shown, even if a grid cell has an extremely high risk or uncertainty value, as long as it is outside the "unreachable area" or "return-to-home critical area" under the current power level, the grid will be marked as gray and off-limits, fundamentally ensuring the physical feasibility and flight safety of the planned path. Safety redundancy power. The settings further consider the additional energy consumption requirements under sudden operating conditions such as headwind flight, battery aging, or emergency obstacle avoidance, and can usually be set to 10% to 20% of the rated power.

[0043] Within the set of feasible candidate grids generated through security filtering, the system further calculates the inspection priority for each grid cell. Specifically, the inspection priority is determined by a weighted sum of the grid cell's risk value, uncertainty value, and degree of inadequate inspection coverage, calculated using the following formula:

[0044] , in, Indicates the first Inspection priority of each grid cell; Indicates the risk value of water and salt stress; Indicates the value of uncertainty; Indicates the first The degree of inspection coverage of a grid cell or its neighborhood; Indicates the degree of inadequacy in inspection coverage; , and These are the weighting coefficients, and Risk value weight The primary weight is determined by risk value; drones prioritize access to grid cells with high inspection priority by using risk value as the primary factor and uncertainty value and the degree of insufficient inspection coverage as secondary factors.

[0045] In this priority evaluation system, Setting it as the dominant weight reflects that the primary goal of inspecting saline-alkali paddy fields is to "identify risks" rather than simply "complete the map." For example, The value can be set to range from 0.5 to 0.7, while and The values ​​are then adjusted between 0.1 and 0.3. This weighting configuration ensures that the drone always prioritizes high-risk areas, avoiding insufficient monitoring of known high-risk areas due to an excessive pursuit of information gain or coverage. Simultaneously, it introduces the degree of inadequate inspection coverage. As a supplementary measure, it can effectively prevent drones from getting stuck in a vicious cycle of local high-risk patches, and encourage them to appropriately expand to surrounding areas that are not fully covered after completing the verification of key areas, so as to take into account the overall situational awareness while ensuring the core objectives. The calculation not only considers whether the current grid has been visited, but also incorporates the attenuation weighting of the coverage contribution of the neighboring grid by the sensor's field of view, making the path planning more consistent with the actual observation geometry.

[0046] Furthermore, to achieve better sequence decision-making in complex dynamic environments, this application incorporates the aforementioned priority evaluation as prior knowledge into a reinforcement learning model, guiding the UAV to autonomously learn the optimal inspection strategy through a reward function. Specifically, the inspection priority is determined through path planning using a reinforcement learning model, whose reward function is:

[0047] , in, express Momentary reward value; Indicates information gain reward; Indicates a reward for coverage of areas with high water and salt stress risk; Indicates a reward for accessing the new region; This indicates a reward for highlighting the value of key sampling points. Indicates path length penalty; Indicates energy consumption penalty; Indicates a penalty for repeated access; This indicates safety penalties such as crossing the boundary, entering a no-fly zone, insufficient battery power, or inability to return to base. , , , , , , and These are the weighting coefficients.

[0048] In this reward function, each reward and penalty has a clear physical mapping relationship and corresponds intrinsically to the aforementioned priority formula. Specifically, information gain reward... When a drone visits a previously unvisited grid cell, the uncertainty value of that grid cell is considered. The magnitude of the uncertainty determines the reward amount; the higher the uncertainty, the greater the reward. This enables reinforcement learning models to automatically learn behavioral patterns of "exploring unknown and ambiguous regions," corresponding to the priority formula. The guiding role of the project. Incentives for coverage of high-risk areas. Risk value of grid cells accessed by drones Positive rewards are provided when the risk level exceeds a preset high-risk threshold, with higher risk values ​​resulting in larger rewards; this directly enhances the model's sensitivity to high-risk targets, corresponding to the priority formula. Dominant position of the item. New area access rewards Provide a fixed positive incentive on each first visit to the grid, along with a path length penalty. and energy consumption penalty This encourages the model to choose compact and efficient paths when exploring new regions, rather than meaningless long jumps. Repeat visit penalty. The system dynamically increases based on historical visit counts, effectively reducing redundant inspections of low-risk areas. (Security penalties) It provides significant negative feedback when a flight crosses a boundary, enters a no-fly zone, or triggers a return-to-base condition, serving as a second line of defense in addition to hard constraints.

[0049] State space during reinforcement learning training Includes the drone's current location Remaining battery power Current heading Visited areas Local risk value Local uncertainty value Return distance and environmental constraints This includes multi-dimensional information. It's worth noting that this application does not allow the model to blindly trial and error from scratch, but rather utilizes a priority formula. Generated prior knowledge is used to accelerate convergence. Specific implementation methods include: in the early stages of training, […]. This approach involves either superimposing a heuristic potential function onto the reward signal or prioritizing the sampling of transfer samples from high-priority regions in the experience replay pool. This is akin to "distilling" the inspection logic of human experts into the model, enabling it to quickly skip the ineffective random exploration phase and directly perform refined optimization on a reasonable policy baseline. The resulting model retains the rigid guarantees of rule-based methods for safety and core objectives while possessing the flexibility of learning methods in handling nonlinear environmental interactions and long-term benefit trade-offs. It maximizes the effectiveness of risk assessment and information acquisition per unit of electricity while ensuring a safe return.

[0050] This embodiment details the scoring and spatially dispersed screening method for key verification sampling points. This method aims to address the technical problem of efficiently transforming risk information acquired through UAV aerial inspections into spatially representative sampling tasks that can be performed by ground personnel. Specifically, recommended key verification sampling points include: calculating the verification priority of candidate sampling points by assigning positive weights to the water-salt stress risk value, uncertainty value, and degree of insufficient inspection coverage of each candidate point, followed by a weighted summation; using the spatial distance between candidate sampling points and existing sampling points, selected verification points, or inspected paths as the minimum spatial interval constraint for subsequent screening stages; sorting all candidate sampling points from highest to lowest verification priority and sequentially screening them according to the minimum spatial interval constraint to output several spatially dispersed key verification sampling points.

[0051] To quantify the aforementioned review priorities, this application provides a comprehensive scoring model. The review priority of candidate sampling points is calculated comprehensively based on the water and salt stress risk value, uncertainty value, degree of insufficient inspection coverage, spatial dispersion distance between the candidate sampling point and existing sampling points or selected review points, and reachability distance to the active inspection path. The calculation formula is as follows:

[0052] , in, Indicates the first The review priority of each candidate sampling point; This indicates the water and salt stress risk value at that point; This represents the uncertainty value at that point; This indicates the level of inspection coverage in the area where the point is located; Indicates the degree of inadequacy in inspection coverage; This represents the normalized spatial dispersion distance between this point and existing sampling points or selected verification points; This represents the normalized reachability distance between the point and the active inspection path or the nearest reachable waypoint; , , , and These are the weighting coefficients.

[0053] In this scoring formula, the design of each parameter embodies a clear agronomic sampling logic and engineering efficiency considerations. The first three items... This constitutes the "value basis" for sampling, ensuring that recommended sampling points have high-risk verification needs in agronomy, high uncertainty clarification needs in cognition, or data gap-filling needs. This is inherently consistent with the inspection priority evaluation logic in the aforementioned embodiments. However, relying solely on the value basis for sampling point recommendation often leads to excessive clustering of sampling points within local high-risk patches, resulting in sample redundancy and a lack of global representativeness. To address this, this application innovatively introduces spatial dispersion distance. As a positive incentive. The higher the value, the farther the candidate point is from existing or selected sampling points, and the higher its score. This design mathematically encourages the system to select points that can fill spatial gaps and improve the uniformity of sample spatial distribution, effectively avoiding the recommendation of multiple functionally overlapping sampling points within the same salt patch.

[0054] At the same time, the formula introduces the reachable distance. As a negative penalty item. The higher the value, the farther the candidate point is from the current active inspection path or the nearest reachable waypoint, and the lower its score. This design fully considers the actual cost and convenience of ground sampling operations. In saline-alkali paddy field environments, areas far from inspection paths or field ridges are often difficult to reach, potentially requiring wading or traversing muddy areas. By incorporating accessibility into the scoring, the system can prioritize recommending points that are both of significant verification value and easily accessible to ground personnel, thereby improving overall operational efficiency, while ensuring sampling value. It should be understood that although this embodiment provides a comprehensive scoring formula containing five factors, in other embodiments, it can be simplified to include only a combination of risk value and spatial dispersion distance, or other evaluation factors such as soil type variability can be added, as long as the core logic takes into account both sampling value and spatial distribution characteristics, it should be considered within the scope of protection of this application.

[0055] In the specific screening process, this application adopts a two-step strategy of "scoring and ranking + spatial constraint filtering," rather than simply weighting and summing all factors and directly selecting the top N points. The specific process is as follows: First, all potential candidate points within the target area are calculated using the above formula. The values ​​are calculated and sorted in descending order to generate a candidate list. Then, a set of selected sampling points is initialized, and the candidate list is iterated through sequentially. For each candidate point in the list, its spatial distance to any point in the set of selected sampling points is checked to see if it meets the preset minimum spatial interval constraint. If the constraint is met, the point is added to the selected set and used as a key verification sampling point in the final output; otherwise, it is not selected. If the value is very high, it is skipped or downgraded, and the next candidate point is checked. This greedy selection mechanism ensures that the final output set of sampling points not only has high priority individually, but also meets the minimum spatial interval requirement as a whole, fundamentally guaranteeing the spatial representativeness of the sampling scheme.

[0056] An example illustrating the technical effectiveness of this mechanism: Suppose there exists a locally heavily salted area measuring 50 meters × 50 meters in a saline-alkali paddy field, with its internal grid... The values ​​are all extremely high. If sorted only by risk value, the system might recommend five adjacent sampling points within the salt patch, leading to repeated sampling by ground personnel within the same small area, while ignoring monitoring of the surrounding transition zone. Using the method of this embodiment, because... The existence of a term means that once the first highest score point located at the center of the salt patch is selected, its neighboring points... The value will decrease sharply, leading to a drop in the overall score; conversely, points located at the edge of the salt patch or in the transition zone far from other selected points, although... Slightly lower, but because Larger sampling points receive higher overall rankings and are selected. The final sampling scheme may include one salt spot center point and four representative points distributed in different surrounding locations, which not only accurately locates the core problem area but also comprehensively grasps the spatial gradient variation pattern of salinization.

[0057] Regarding weighting coefficients , , , and In terms of adjustment strategies, this application provides flexible configuration space to adapt to different application scenarios. For example, in detailed surveys for scientific research purposes, in order to obtain the most detailed spatial variation information, the weight of spatial dispersion distance can be appropriately increased. This forces sampling points to be distributed as evenly as possible throughout the study area; however, in rapid diagnosis aimed at emergency investigation of agricultural production, the weight of reachability distance can be appropriately increased to improve response speed and reduce personnel burden. This concentrates recommended locations in areas with convenient transportation. Furthermore, if the inspection is currently in its initial stages and historical data is scarce, the uncertainty weight can be increased. The focus should be on exploring the unknown; if it is in a long-term monitoring phase, the weight of the risk value can be increased. It focuses on tracking the dynamic changes of known severely affected areas. This parameter adjustability allows the same algorithm framework to adapt to diverse business needs, enhancing the practicality and robustness of the technical solution.

[0058] This embodiment details the method for dynamically updating the water and salt stress risk map based on confidence weighting. This method is a core component in enabling the inspection system to transition from "single-time monitoring" to "adaptive evolution," aiming to address the issues of accuracy differences and cognitive iteration in the fusion of multi-source heterogeneous data. Specifically, updating the water and salt stress risk map based on UAV inspection feedback data and ground-based measured data includes: when converting UAV inspection feedback data into new risk observation values, performing orthorectification, geometric correction, and radiometric correction on the UAV imagery; extracting low-altitude vegetation index, moisture index, salinity-related index, or texture features at a spatial scale consistent with the grid environment; and obtaining the UAV-observed risk value according to the same risk factor transformation rules as the initial risk map.

[0059] In this process, the rigor of data preprocessing directly determines the accuracy of the update. Orthorectification is used to eliminate image distortion caused by UAV flight attitude, geometric correction ensures precise registration of UAV imagery with satellite base maps and grid coordinate systems, and radiometric correction eliminates the interference of changes in illumination conditions on spectral characteristics. More critically, there is the constraint of "spatial scale consistent with the grid environment," which means resampling or aggregating the high-resolution images acquired by the UAV to a size that perfectly matches the defined grid cells. For example, if the grid is 20m × 20m, then the UAV images with a 5cm resolution need to be aggregated into the statistical feature values ​​within that grid. Only by achieving this strict spatial alignment can the risk value of each grid cell be corrected one by one, avoiding erroneous updates caused by spatial misalignment. Simultaneously, "following the same risk factor transformation rules as the initial risk map" ensures semantic consistency between the old and new data. That is, whether it is satellite data or UAV data, the mathematical logic for converting it into risk factors (such as normalization direction and threshold setting) remains consistent, making risk values ​​from different sources comparable and fusionable.

[0060] Furthermore, for the processing of ground-measured data, the conductivity, pH, soil moisture content, or water conductivity in the ground verification data are normalized to obtain the ground-observed risk values. Although ground-measured data has high accuracy, it is usually discrete point data with different dimensions (e.g., conductivity is measured in dS / m, pH is a dimensionless value). Therefore, it needs to be converted into 0-1 risk observation values ​​with the same dimensions as the risk map. As one implementation method, piecewise linear mapping or membership functions can be used for normalization. For example, for soil conductivity EC, a threshold for mild salinity can be set. and the threshold for severe salt damage ,when When the risk value is 0, When the risk value is 1, linear interpolation occurs when the value falls between 1 and 1. This approach abstracts physical measurements into standardized risk semantics, allowing them to be seamlessly integrated into the numerical system of risk maps.

[0061] Based on this, this application proposes a confidence-weighted update mechanism. When both UAV observation risk values ​​and ground observation risk values ​​exist simultaneously, a new risk observation value is obtained by weighting them according to their respective confidence levels. The water and salt stress risk map is then updated using a confidence-weighted update method.

[0062] , , in, This indicates the updated risk value for water and salt stress. This indicates the water and salt stress risk value before the update; This indicates a new risk observation obtained from drone inspection data, ground verification data, or a confidence-weighted average of both. This indicates an update to the weights; Indicates the confidence level of the new observation data; This indicates the confidence level of the original risk map.

[0063] This update formula embodies the idea of ​​Bayesian estimation, its core being the dynamic adjustment of the contribution ratio between new and old information based on confidence levels, rather than simply overwriting old data with new data. Direct overwriting is not possible because the reliability of different data sources varies significantly, and historically accumulated prior knowledge may be more stable in certain situations than a single observation. The update weights in the formula... It is a dynamic variable: the confidence level of new observation data Much higher than the confidence level of the original risk map hour, Approaching 1, at this point The system fully trusts new data and quickly corrects its understanding; conversely, if the new data has low confidence (e.g., drone imagery is obscured by clouds or ground sampling points are not representative enough), while the original risk map has accumulated high confidence based on long-term monitoring, then... As the value approaches zero, the system tends to retain historical priors, avoiding drastic fluctuations in the risk map due to a single low-quality observation. This mechanism endows the system with the ability to "learn prudently," maintaining both sensitivity to external changes and stability of internal cognition.

[0064] Regarding confidence level The assignment principle in this application follows the logic of "true value is superior to inversion, and near field is superior to far field." Generally speaking, ground-based measured data, as directly measured true values, has the highest confidence level; UAV low-altitude remote sensing data has high resolution and strong timeliness, resulting in the next highest confidence level; satellite remote sensing data, although with wide coverage, is easily affected by the atmosphere and has lower resolution, resulting in the lowest relative confidence level. In specific implementation, benchmark confidence scores can be preset for different data sources, such as ground-based measured data... UAV observation Satellite inversion Furthermore, the confidence level can be dynamically adjusted based on data quality indicators. For example, the higher the cloud mask ratio and the lower the signal-to-noise ratio of UAV imagery, the lower its actual confidence level should be. When the same grid contains both new UAV and ground-based observations, their respective confidence levels can be used to merge them into a single comprehensive new risk observation. and its corresponding overall confidence level Then substitute the above formula into the updated risk map.

[0065] Finally, the updated water and salt stress risk map It serves not only as the output of current monitoring results, but more importantly, as the basis for planning the next round of inspection routes. Prior input. As the number of inspection rounds increases, the confidence level in high-risk areas gradually accumulates, and the uncertainty gradually decreases. The inspection strategy of drones will gradually converge from the initial extensive exploration to precise monitoring of specific stubborn risk points. This data-driven closed-loop iterative mechanism enables the inspection system to continuously self-optimize along with the evolution of the farmland environment, truly realizing the intelligent and long-term management of water and salt stress in saline-alkali paddy fields.

[0066] In this embodiment, the UAV active inspection path planning method for saline-alkali paddy fields provided in the aforementioned embodiments is applied to a typical saline-alkali rice planting area in northern my country to verify the effectiveness and superiority of the technical solution in a real agricultural production environment. The experimental area covers approximately 25 square kilometers, with flat terrain but significant spatial heterogeneity in soil salinization, making it an ideal scenario for testing active inspection strategies.

[0067] In the specific implementation, the target area was first discretized. Considering both the swath width of the UAV sensor and the average scale of the saline-alkali patches, the study area was divided into a regular grid environment of 50 rows × 50 columns, with each grid cell approximately 100 meters × 100 meters in size. The inspection platform used was a quadcopter UAV with an endurance of approximately 30 minutes and a payload including a multispectral camera and a visible light camera. When constructing the initial water and salt stress risk map, Sentinel-2 multispectral imagery, a digital elevation model, and historical soil survey data were integrated. Given that salinity stress is the primary limiting factor in this area, the weight parameters of each evaluation model were specifically configured. Specifically, risk value weights were set when calculating the inspection priority. The uncertainty value weight is 0.6. The weight for insufficient inspection coverage is 0.3. The weighting is 0.1. This risk-driven weighting configuration ensures that drones, with limited power, always prioritize monitoring needs in high-risk areas. When recommending key verification sampling points, to balance sample representativeness and ground accessibility, the risk value weight m in the verification priority formula is set to 0.4, the uncertainty weight n to 0.2, the insufficient coverage weight q to 0.1, the spatial dispersion distance weight r to 0.2, and the accessibility distance weight s to 0.1. The higher spatial dispersion distance weight effectively avoids excessive clustering of sampling points within local heavy salt patches, while the introduction of the accessibility distance weight significantly improves the efficiency of ground personnel performing sampling tasks.

[0068] During a complete inspection mission, the system first generates an inspection strategy based on an initial risk map and uncertainty map. The reinforcement learning model employs a deep Q-network algorithm, deployed on an airborne computing unit or ground station after offline training. The UAV departs from a preset takeoff point, continuously sensing the local risk status and remaining battery power at its current location, dynamically selecting the next hop grid. During flight, airborne sensors simultaneously collect high-resolution remote sensing data. After the mission, the system automatically selects eight key verification sampling points based on the actual flight trajectory and collected data and outputs them to the ground terminal. Subsequently, using UAV orthophotos and some ground-measured conductivity data, the initial risk map is corrected through a confidence-weighted update mechanism, generating a more accurate new prior map, completing the full closed loop from perception to update.

[0069] To visually demonstrate the technical effects of the method in this application, such as Figure 5 As shown, the active inspection path generated in this embodiment is compared and analyzed with the traditional full-coverage serpentine path. Figure 5The left side shows a traditional full-coverage route, characterized by mechanically traversing every grid cell, allocating the same inspection resources to both low-risk healthy rice paddies and high-risk saline-alkali patches. While this method guarantees theoretical full coverage, in practice it results in a significant amount of wasted flight time and often fails to complete a detailed scan of the entire area before power depletion, leading to the omission of some high-risk peripheral areas. In contrast, Figure 5 The active inspection path shown on the right exhibits significant intelligent features. After takeoff, the drone flies directly to the high-risk area A (filled with diagonal lines) in the image for detailed, circling observation. It then quickly jumps to high-risk areas B and C. Along the way, it only performs edge detection on the high-uncertainty areas D and E (shown by dashed boxes) to eliminate cognitive ambiguity. Large areas of low-risk healthy regions are either skipped or only sparsely inspected. This non-uniform path distribution pattern reflects the algorithm's precise trade-off between "risk value" and "information gain."

[0070] Quantitative comparison results show that, under the same battery capacity and flight time constraints, the total flight distance of the active inspection path using the method of this application is reduced by approximately 35% compared to the traditional full-coverage path, while the number of visits and dwell time in high-risk areas increases by approximately 45%. More importantly, through the spatially dispersed screening mechanism of key verification sampling points, the recommended ground sampling points cover all major salinization gradient zones in the experimental area, significantly improving the spatial representativeness of the samples. The risk map updated based on the data feedback from these high-quality sampling points shows a more than 20% higher degree of agreement with the actual salinity distribution compared to the initial risk map. This result fully demonstrates that the method proposed in this application can not only significantly improve the operational efficiency of UAV inspection and reduce energy waste, but also continuously improve the accuracy and timeliness of water and salt stress monitoring in saline-alkali paddy fields through a closed-loop mechanism of "aerial screening-ground verification-model iteration," providing reliable data support and technical means for precision agricultural management. It should be understood that the above parameter values ​​and comparative data are only for illustrating the implementation effect of the technical solution of this application in a specific scenario and are not intended to limit the scope of protection of this application. In practical applications, relevant personnel can make adaptive adjustments to the above weight coefficients and thresholds according to specific field conditions, crop varieties and management objectives. As long as they follow the core ideas of proactive inspection and closed-loop updating disclosed in this application, they should be considered to fall within the protection scope of this application.

[0071] The above inventions are merely a few specific embodiments of the present invention. However, the embodiments of the present invention are not limited thereto, and any variations that can be conceived by those skilled in the art should fall within the protection scope of the present invention.

Claims

1. A method for active inspection path planning by unmanned aerial vehicles (UAVs) in saline-alkali paddy fields, characterized in that, include: Identify the target saline-alkali paddy field area; The target saline-alkali paddy field area is discretized into a grid environment, and multi-source data in different grid environments are obtained; Based on the multi-source data, water and salt stress-related characteristic indicators were extracted, and the water and salt stress risk values ​​and uncertainty values ​​of different grids were calculated. The water and salt stress risk values ​​and uncertainty values ​​are mapped to the corresponding grid environment to obtain the water and salt stress risk map and uncertainty map; The drone plans its inspection path based on the water and salt stress risk map and uncertainty map; Based on the coverage of the inspection route and the water and salt stress risk map, it is recommended to focus on verifying the sampling points; The water and salt stress risk map and uncertainty map are updated based on the feedback data from drone inspections and ground measurement data.

2. The method for active inspection path planning of saline-alkali paddy fields by unmanned aerial vehicles as described in claim 1, characterized in that, When the UAV plans its inspection path based on the risk map and the uncertainty map, the inspection priority of a candidate grid cell that meets the safe return-to-home condition is determined by a weighted sum of the grid cell's risk value, uncertainty value, and degree of insufficient inspection coverage. The calculation formula is as follows: ， in, Indicates the first Inspection priority of each grid cell; Indicates the risk value of water and salt stress; Indicates the value of uncertainty; Indicates the first The degree of inspection coverage of a grid cell or its neighborhood; Indicates the degree of inadequacy in inspection coverage; , and These are the weighting coefficients, and Risk value weight The drone prioritizes access to grid cells with high inspection priority, using risk value as the primary factor and uncertainty value and the degree of insufficient inspection coverage as secondary factors.

3. The method for active inspection path planning of saline-alkali paddy fields by unmanned aerial vehicles as described in claim 2, characterized in that, The safe return-to-home condition is: the drone's current remaining battery power is greater than the sum of the battery power required to return to the return point after traveling from the current location to the candidate grid cell and the safety redundancy battery power; candidate grid cells that do not meet the safe return-to-home condition are not included in the calculation of the inspection priority.

4. The method for active inspection path planning of saline-alkali paddy fields by unmanned aerial vehicles as described in claim 2, characterized in that, The inspection priority is determined through a reinforcement learning model for path planning. The reward function of the reinforcement learning model is: ， in, This represents the reward value at time t; Indicates information gain reward; This indicates a reward for coverage of areas with high water and salt stress risk; This indicates a reward for accessing new areas; This indicates a reward for highlighting the value of key sampling points. Indicates path length penalty; Indicates energy consumption penalty; Indicates a penalty for repeated access; This indicates safety penalties such as crossing the boundary, entering a no-fly zone, insufficient battery power, or inability to return to base. , , , , , , and These are the weighting coefficients; In the reward function, the information gain reward is determined based on the uncertainty value of a previously unvisited grid cell when the drone visits it; the higher the uncertainty, the greater the reward. The high-risk area coverage reward provides a positive reward when the risk value of a grid cell visited by the drone exceeds a preset high-risk threshold; the higher the risk value, the greater the reward. The new area access reward provides a fixed positive reward each time the drone visits a grid cell for the first time. The path length penalty and the energy consumption penalty are determined based on the distance the drone travels in a single step and its energy consumption. The repeated access penalty is determined based on the historical number of visits to the current grid cell; the more times it is accessed, the greater the penalty. The safety penalty is triggered when the drone crosses the boundary, enters a no-fly zone, or has insufficient remaining battery power to safely return to base. The drone selects an action to perform based on its current state using the reinforcement learning model. After performing the action, it receives a reward and moves to the next state. The path planning model is trained by maximizing the accumulated expected reward.

5. The method for active inspection path planning of saline-alkali paddy fields by unmanned aerial vehicles as described in claim 1, characterized in that, The formula for calculating the water-salt stress risk value is as follows: ， in, Indicates the first Water and salt stress risk value for each grid cell; Indicates the salt risk factor; Indicates vegetation stress risk factors; Indicates the risk factors for abnormal moisture levels; Indicates the risk factor for heat stress; Indicates a priori risk factors for saline-alkali land or terrain; , , , and Let represent the weight coefficients of each factor, and + + + + =1.

6. The method for active inspection path planning of saline-alkali paddy fields by unmanned aerial vehicles as described in claim 5, characterized in that, The formula for calculating the uncertainty value is as follows: ， in, Let be the uncertainty value of the i-th grid cell. Indicates standard deviation, Normalization is indicated; the uncertainty value represents the degree of consistency of the judgment results of each risk factor for the same grid cell. The larger the value, the higher the dispersion between each risk factor and the more unstable the remote sensing judgment.

7. The method for active inspection path planning of saline-alkali paddy fields by unmanned aerial vehicles as described in claim 1, characterized in that, The drone will automatically return to home when its remaining battery power is less than or equal to the sum of the battery power required to return to the home point from its current location and the safety margin battery power. ， in, This is the current remaining battery level; For the drone from its current location The amount of electricity required to return to the starting point; For safety, redundant power supply.

8. The method for active inspection path planning of saline-alkali paddy fields by unmanned aerial vehicles as described in claim 1, characterized in that, The recommended key review sampling points include: The review priority of candidate sampling points is calculated by weighted summation after assigning positive weights to the water and salt stress risk value, uncertainty value and the degree of insufficient inspection coverage of the candidate point; the spatial distance between the candidate sampling point and existing sampling points, selected review points or inspected paths is used as the minimum spatial interval constraint for the subsequent screening stage; all candidate sampling points are sorted from high to low review priority and screened in sequence according to the minimum spatial interval constraint to output a number of spatially dispersed key review sampling points.

9. The method for active inspection path planning of saline-alkali paddy fields by unmanned aerial vehicles as described in claim 8, characterized in that, The review priority of the candidate sampling points is calculated comprehensively based on the water and salt stress risk value, uncertainty value, degree of insufficient inspection coverage, spatial dispersion distance between the candidate sampling points and existing sampling points or selected review points, and reachability distance from the active inspection path. The calculation formula is as follows: ， in, Indicates the first The review priority of each candidate sampling point; This indicates the water and salt stress risk value at that point; This represents the uncertainty value at that point; This indicates the level of inspection coverage in the area where the point is located; Indicates the degree of inadequacy in inspection coverage; This represents the normalized spatial dispersion distance between this point and existing sampling points or selected verification points; This represents the normalized reachability distance between the point and the active inspection path or the nearest reachable waypoint; , , , and These are the weighting coefficients.

10. The method for active inspection path planning of saline-alkali paddy fields by unmanned aerial vehicles as described in claim 1, characterized in that, The process of updating the water and salt stress risk map based on UAV inspection feedback data and ground-based measured data includes: when converting UAV inspection feedback data into new risk observation values, performing orthorectification, geometric correction, and radiometric correction on the UAV imagery, and extracting low-altitude vegetation index, moisture index, salinity-related index, or texture features at a spatial scale consistent with the grid environment, and obtaining UAV observation risk values ​​according to the same risk factor conversion rules as the initial risk map; normalizing the conductivity, pH, soil moisture content, or water conductivity in the ground verification data to obtain ground observation risk values; when both UAV observation risk values ​​and ground observation risk values ​​exist simultaneously, obtaining new risk observation values ​​by weighting their respective confidence levels, and updating the water and salt stress risk map using a confidence-weighted update method. ， in, This indicates the updated risk value for water and salt stress. This indicates the water and salt stress risk value before the update; This indicates a new risk observation obtained from drone inspection data, ground verification data, or a confidence-weighted average of both. This indicates an update to the weights; Indicates the confidence level of the new observation data; This indicates the confidence level of the original risk map.

Images