A dynamic patrol method and device of an unmanned aerial vehicle on farmland

Abstract

This invention discloses a method and apparatus for dynamic inspection of farmland using unmanned aerial vehicles (UAVs). The invention relates to the technical field of dynamic inspection methods. It involves acquiring multiple inspection images, determining multiple pest and disease characteristics of rice based on these images and the morphology of the rice, and determining the damage level of rice in a key inspection area based on the characteristic location, corresponding morphological features, and distribution location of invasive alien species for each pest and disease characteristic, thus improving the accuracy of rice damage level assessment in the key inspection area. Therefore, multiple abnormal rice areas are identified based on the recognition of dynamic distribution maps. Abnormal inspection events of the UAV are determined based on the regional locations of these abnormal rice areas, the associated water flow areas, and the current location of the UAV. Dynamic inspection events of the UAV are then determined based on these abnormal inspection events, the weather conditions corresponding to each abnormal rice area and the farmland, further improving the accuracy of dynamic inspection events.

Description

Technical Field

[0001] This invention relates to the technical field of dynamic inspection methods, and more particularly to a method and apparatus for dynamic inspection of farmland by unmanned aerial vehicles (UAVs). Background Technology

[0002] With the development of technology, farmland refers to land that has been cultivated and prepared by humans for agricultural production. In this context, farmland includes infrastructure such as field ridges, irrigation canals, drainage ditches, and farm roads. Drones fly over farmland to inspect it. Current technology collects images of the farmland by tracking the drone's inspection route. However, this method neglects the characteristic locations of various pests and diseases, as well as the distribution of invasive species, affecting the accuracy of assessing the damage level of rice in key inspection areas. This results in low accuracy for dynamic drone inspections. Summary of the Invention

[0003] The purpose of this invention is to overcome the shortcomings of the prior art. This invention provides a method and device for dynamic inspection of farmland by unmanned aerial vehicles (UAVs).

[0004] This invention provides a method for dynamic inspection of farmland using unmanned aerial vehicles (UAVs), comprising:

[0005] Based on the initial distribution map of the farmland and the current location of the drone, the inspection path of the drone for the farmland is determined. Based on the inspection path and the distribution pattern of rice, multiple key inspection nodes are identified. The multiple key inspection nodes are located at different altitudes.

[0006] The key inspection area is determined based on the node location and the corresponding node influence range of each key inspection node. Sub-inspection routes are then determined based on the key inspection area, the corresponding rice paddies, and the drones.

[0007] The drone flies along the sub-inspection route and collects multiple inspection images. Based on the multiple inspection images and the morphology of the rice, it determines multiple pest and disease characteristics of the rice. Based on the characteristic location of each pest and disease characteristic, the corresponding characteristic morphology, and the distribution location of invasive alien species, it determines the damage level of the rice in the key inspection area.

[0008] Based on the damage level of rice in each key inspection area, the corresponding regional location, and multiple inspection images, a damage heat map of the key inspection area is determined. Based on each damage heat map, the initial distribution map of farmland, and the regional morphology of the key inspection area, a dynamic distribution map of farmland is determined.

[0009] Multiple abnormal rice areas are identified based on the dynamic distribution map. Abnormal patrol events of the drone are determined based on the regional location of the multiple abnormal rice areas, the water flow areas associated with the abnormal rice areas, and the current location of the drone. Dynamic patrol events of the drone are determined based on the abnormal patrol events of the drone, the weather conditions corresponding to each abnormal rice area and farmland.

[0010] This invention provides a device for dynamic inspection of farmland using unmanned aerial vehicles (UAVs). The device is applied to the aforementioned method for dynamic inspection of farmland using UAVs. The device includes:

[0011] The key inspection node module is used to determine the inspection path of the UAV on the farmland based on the initial distribution map of the farmland and the current position of the UAV. Based on the inspection path and the distribution pattern of rice, multiple key inspection nodes are determined; the multiple key inspection nodes are located at different heights.

[0012] The sub-inspection route module is used to determine the key inspection area based on the node location and the corresponding node influence range of each key inspection node, and to determine the sub-inspection route based on the key inspection area, the corresponding rice and drones.

[0013] The rice damage level module is used by drones to fly along sub-inspection routes and collect multiple inspection images. Based on the multiple inspection images and the morphology of rice, multiple pest and disease characteristics of rice are determined. Based on the characteristic location of each pest and disease characteristic, the corresponding characteristic morphology, and the distribution location of invasive alien species, the damage level of rice in the key inspection area is determined.

[0014] The dynamic distribution map module is used to determine the damage heat map of each key inspection area based on the rice damage level, corresponding regional location, and multiple inspection images. Based on each damage heat map, the initial distribution map of farmland, and the regional morphology of the key inspection area, the dynamic distribution map of farmland is determined.

[0015] The dynamic patrol event module is used to identify multiple abnormal rice areas based on the dynamic distribution map. It determines the abnormal patrol events of the drone based on the regional location of multiple abnormal rice areas, the water flow areas associated with the abnormal rice areas, and the current location of the drone. It determines the dynamic patrol events of the drone based on the abnormal patrol events of the drone, the weather conditions corresponding to each abnormal rice area and farmland.

[0016] Compared with the prior art, the beneficial effects of the present invention are:

[0017] In this embodiment of the invention, the method determines the inspection path of the UAV based on the initial distribution map of the farmland and the current position of the UAV. Based on this inspection path and the distribution pattern of rice, multiple key inspection nodes are identified. These key inspection nodes are located at different altitudes. A key inspection area is determined based on the node position and corresponding influence range of each key inspection node. A sub-inspection route is determined based on this key inspection area, the corresponding rice, and the UAV. The UAV flies along the sub-inspection route and collects multiple inspection images. Based on these images and the morphology of the rice, multiple pest and disease characteristics of the rice are determined. The damage level of the rice in the key inspection area is determined based on the characteristic position, corresponding morphology, and distribution location of each pest and disease characteristic. The introduction of multiple key inspection nodes further controls the multiple pest and disease characteristics of rice, incorporating a holistic consideration of the characteristic position, corresponding morphology, and distribution location of each pest and disease characteristic, thus improving the accuracy of the rice damage level in the key inspection area.

[0018] Therefore, based on the damage level of rice in each key inspection area, the corresponding regional location, and multiple inspection images, a damage heat map of that key inspection area is determined. Based on these damage heat maps, the initial distribution map of farmland, and the regional morphology of the key inspection areas, a dynamic distribution map of farmland is determined. Multiple abnormal rice areas are identified based on the dynamic distribution map. Abnormal drone inspection events are determined based on the regional location of these abnormal rice areas, the associated water flow areas, and the current location of the drone. Dynamic drone inspection events are then determined based on these abnormal inspection events, the weather conditions corresponding to each abnormal rice area and farmland. The introduction of a dynamic farmland distribution map further controls abnormal drone inspection events, achieving a holistic consideration of abnormal drone inspection events, abnormal rice areas, and the corresponding weather conditions, thus improving the accuracy of dynamic drone inspection events. Attached Figure Description

[0019] Figure 1 This is a flowchart illustrating the dynamic inspection method of farmland by drone in an embodiment of the present invention.

[0020] Figure 2 This is a flowchart illustrating step S11 of the method for dynamic inspection of farmland by drone in an embodiment of the present invention.

[0021] Figure 3 This is a flowchart illustrating step S12 of the method for dynamic inspection of farmland by drone in an embodiment of the present invention.

[0022] Figure 4 This is a flowchart illustrating step S13 of the method for dynamic inspection of farmland by drone in an embodiment of the present invention.

[0023] Figure 5 This is a flowchart illustrating step S14 of the method for dynamic inspection of farmland by drone in an embodiment of the present invention.

[0024] Figure 6 This is a flowchart illustrating step S15 of the method for dynamic inspection of farmland by drone in an embodiment of the present invention.

[0025] Figure 7 This is a schematic diagram of the structural composition of the drone-based dynamic inspection device for farmland in an embodiment of the present invention. Detailed Implementation

[0026] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention.

[0027] Please see Figures 1 to 7 A method for dynamic inspection of farmland using drones, applied to dynamic inspection scenarios; the method for dynamic inspection of farmland using drones includes:

[0028] Step S11: Determine the inspection path of the UAV for the farmland based on the initial distribution map of the farmland and the current position of the UAV. Based on the inspection path and the distribution pattern of rice, determine multiple key inspection nodes; the multiple key inspection nodes are located at different heights.

[0029] Step S12: Determine the key inspection area based on the node location and corresponding node influence range of each key inspection node, and determine the sub-inspection route based on the key inspection area, the corresponding rice and drones;

[0030] Step S13: The drone flies along the sub-inspection route and collects multiple inspection images. Based on the multiple inspection images and the morphology of the rice, it determines multiple pest and disease characteristics of the rice. Based on the characteristic location of each pest and disease characteristic, the corresponding characteristic morphology, and the distribution location of invasive alien species, it determines the damage level of the rice in the key inspection area.

[0031] Step S14: Determine the damage heat map of each key inspection area based on the rice damage level, corresponding regional location, and multiple inspection images. Determine the dynamic distribution map of the farmland based on each damage heat map, the initial distribution map of the farmland, and the regional morphology of the key inspection area.

[0032] Step S15: Based on the identification of the dynamic distribution map, identify multiple abnormal rice areas, determine the abnormal patrol events of the drone based on the regional location of the multiple abnormal rice areas, the water flow areas associated with the abnormal rice areas and the current location of the drone, and determine the dynamic patrol events of the drone based on the abnormal patrol events of the drone, the weather conditions corresponding to each abnormal rice area and farmland.

[0033] refer to Figure 2 In step S11, the specific steps are as follows:

[0034] S111: Locate the farmland and collect an initial distribution map of the farmland; determine the drone's flight area based on the farmland's location, the corresponding initial distribution map, and the drone's current location; determine the drone's inspection path for the farmland based on the drone's flight area and the farmland's distribution pattern.

[0035] S112: Based on the detection of the inspection path, multiple sub-inspection paths are determined. At the same time, the distribution style of rice in the farmland is collected. Multiple inspection nodes are determined according to the multiple sub-inspection paths, the position difference between two adjacent sub-inspection paths, and the distribution style of rice.

[0036] S113: Based on the node positions and corresponding node priorities of multiple inspection nodes, multiple key inspection nodes are determined. These multiple key inspection nodes are located in different positions in the farmland, covering both the internal and upper spaces of the farmland, and are located at different heights.

[0037] In the embodiments of this application, the geographical coordinates of farmland are obtained through a Global Navigation Satellite System (GNSS, such as GPS, BeiDou), which is typically a polygonal region defined by multiple vertices; secondly, the "initial distribution map" is not a simple satellite image, but a geographic information model fused from multiple sources; it should at least include:

[0038] Digital Elevation Model (DEM) or Digital Surface Model (DSM): Provides information on the topographic relief of farmland, which is crucial for planning safe flight altitudes and preventing drones from crashing into the ground due to terrain changes; Plot boundary vector data: Precisely defines the boundaries of farmland, preventing drones from flying into adjacent plots or no-fly zones; Baseline spectral imagery: Can be recent multispectral satellite imagery, used to initially assess the overall growth distribution of rice, providing prior knowledge for subsequent "distribution style" analysis; Fixed obstacle information: Such as utility poles, pumping stations, shelterbelts, etc. along field ridges, marked on the initial map as hard constraints for path planning.

[0039] An outward buffer operation is performed on the polygonal boundary of the farmland to generate a larger area. The width of the buffer depends on the minimum turning radius of the drone, the error margin of the positioning system, and the edge distortion of the camera's field of view (FOV). For example, setting a buffer of 5-10 meters can ensure that the camera can completely cover the farmland when shooting the boundary, and that the drone has enough space to maneuver.

[0040] Based on DEM data, the system calculates the lower limit of safe flight altitude for each point within the flight area, ensuring that the drone is always a safe distance (e.g., 3-5 meters) above the highest point on the ground (such as field ridges or crop canopies). At the same time, the system takes into account the drone's current location (Home point). If the Home point is outside the expanded flight area, the system will automatically generate a safe access path from the Home point to the entrance of the flight area.

[0041] Analyze the geometry of farmland boundaries; for approximately rectangular, flat farmland, "bow-shaped" or "S-shaped" paths are preferred because these paths have fewer turns and the highest coverage efficiency; for long or irregularly shaped farmland, "spiral" or "sub-region-ox-plowing" hybrid paths based on convex hull decomposition are used to reduce ineffective flights.

[0042] After determining the path pattern, key parameters need to be calculated; the most important is the flight path spacing (Spacing), which is determined by the camera's field of view (FOV), the preset flight altitude, and the required image overlap (usually forward overlap >70%, lateral overlap >50%). The formula can be simplified to: Spacing = Image_Width × (1 - Overlap_Ratio). After generating the initial path, the system will also optimize it; for example, adjusting the path's main axis according to the prevailing wind direction or the direction of sunlight to reduce disturbance to the rice or avoid shadow effects; and fine-tuning the path based on DEM data to maintain a constant relative height to the ground in three-dimensional space.

[0043] Furthermore, the system analyzes the DEM or plot boundary data in the initial distribution map. When the path crosses field ridges, irrigation ditches, or the boundary between plots of different varieties / sowing dates, it automatically segments the path. Each sub-path contains complete information such as the starting point coordinates, the ending point coordinates, the expected flight altitude, and the speed, forming an ordered path queue. Here, "detection" refers to analyzing the attributes of the planned path, such as calculating the length, heading angle, and distance to nearby obstacles of each sub-path, providing spatial location parameters for the generation of subsequent nodes.

[0044] The distribution pattern of rice in the farmland was collected, and the main data sources included: historical remote sensing images (analyzing multispectral satellite or UAV images from a few days before the start of the mission to calculate the spatial distribution map of vegetation indices such as NDVI), agronomic knowledge base (the distribution pattern of rice is also related to its growth stage and varietal characteristics), and real-time preview data (the UAV can conduct a quick, low-resolution overview flight before the formal inspection).

[0045] The system uses image processing algorithms to convert the above data into a quantified "style feature map". For example, it generates a "growth uniformity map". The lower the value in the map, the more uniform the rice growth in the area. The higher the value, the greater the difference in growth (there are problems such as missing seedlings and uneven fertility).

[0046] The system employs two strategies to generate candidate nodes: one is path-based node generation, where the system calculates the spatial relationship between adjacent sub-paths, particularly the "position difference" (i.e., heading spacing and endpoint offset); the middle position of two parallel sub-paths, or at the turning point (waypoint) of the path, are natural candidate node positions, and these nodes ensure coverage of the "blind spots" between paths.

[0047] Node generation based on distribution style: The system performs spatial overlay analysis of the "rice distribution style map" and sub-paths; in areas marked as "non-uniform", "high variability" or "stressed" in the style map, the system generates candidate nodes with a higher density; the final "inspection node" is the weighted fusion result of the above two strategies; the system calculates an "importance score" for each candidate node, for example: score = α × (weight based on path relationship) + β × (weight based on distribution style); in homogeneous regions, the weight of α is higher; in heterogeneous regions, the weight of β is significantly increased.

[0048] Therefore, multiple key inspection nodes are determined based on the node positions and corresponding node priorities of multiple inspection nodes. These key inspection nodes are located in different positions in the farmland, covering both the internal and upper spaces of the farmland. They are also located at different heights, taking into account the overall consideration of the node positions and corresponding node priorities of multiple inspection nodes, thus ensuring the accuracy of the multiple key inspection nodes.

[0049] At this point, the priority of a node is determined by a weighted average of several factors, including: whether the node is located at the geometric center of the farmland, a key intersection, or the connection point between different sub-regions, which are crucial for understanding the overall layout; whether the node is located within the "non-uniform / stressed style" region identified in S112; nodes located at the center of highly variable, low-activity patches should have their priority significantly increased; and whether the node is the starting point, ending point, or key turning point of the path, which are natural anchor points for route execution.

[0050] Are there any pre-set permanent sampling points, historically high-incidence areas of diseases, or sensor installation locations in the farmland? Nodes located near these POIs should be given high priority. The system assigns weights to each factor (e.g., the weight of abnormal correlation is the highest) and calculates a comprehensive score for each inspection node. The final list of "key inspection nodes" is determined by setting a score threshold or by directly selecting the top N nodes.

[0051] The selected set of key nodes must be spatially extensive, not just concentrated in a corner of the farmland. They should collectively form a "sampling network" that reflects the overall condition of the farmland. More importantly, the observation tasks of the nodes are assigned to different spatial levels: Intra-canopy: refers to the space below the rice canopy, close to the ground or water surface; observation of the intra-canopy aims to detect stem diseases, lodging, water surface weeds, or underwater / near-ground organisms such as golden apple snails. Above-canopy: refers to the space above the rice canopy; observation of the above-canopy aims to assess the overall growth, density, color of the canopy, and the presence of large-scale leaf diseases or pests.

[0052] For observation nodes in the "above space": the flight altitude is usually relatively high, such as 25-30 meters; at this altitude, the camera has a large field of view and can acquire macroscopic images of the canopy, which is convenient for calculating vegetation index and assessing overall density; the altitude will be finely adjusted according to the DEM data of S111 to ensure that the relative altitude is constant.

[0053] For observation nodes in the "internal space": the flight altitude is significantly reduced, for example, 5-10 meters; at this altitude, drones equipped with high-resolution cameras can clearly capture the texture and color changes of individual rice paddies, and even small insects. This low-altitude flight places extremely high demands on the drone's altitude hold accuracy and obstacle avoidance capabilities.

[0054] For the highest priority node, the system can even plan a "vertical observation sequence", in which the drone first takes a panoramic image at a height of 30 meters, then descends vertically to a height of 8 meters to take a detailed image, thus achieving correlation analysis from macro to micro.

[0055] refer to Figure 3 In step S12, the specific steps are as follows:

[0056] S121: Among multiple key inspection nodes, the node shape of the key inspection node is determined based on the detection of each key inspection node, and the node influence range of the key inspection node is determined based on the node shape of the key inspection node, the corresponding node position and the shape of the surrounding rice. The key inspection area is determined according to the node position of each key inspection node, the corresponding node influence range and the distribution range of the surrounding rice.

[0057] S122: Based on the detection of key inspection areas, determine the corresponding rice and mark the current form of the rice. Determine the first sub-inspection route based on the key inspection area and the form of the corresponding rice. Determine the second sub-inspection route based on the key inspection area and the flight form of the drone. Determine the sub-inspection route based on the first sub-inspection route, the second sub-inspection route and the current battery level of the drone.

[0058] In the embodiments of this application, among multiple key inspection nodes, the node morphology of the key inspection node is determined based on the detection of each key inspection node. The "node morphology" is a set of information containing multiple dimensions, which determines the observation target of the node and is the core of the morphology. Depending on the task type, the node morphology may include a preset or suggested flight altitude (e.g., 30 meters for macroscopic observation and 8 meters for microscopic observation) and camera attitude (e.g., vertical downward -90°, or tilted -45° to observe the canopy side).

[0059] In addition, the node morphology also defines the required data type and precision; for example, is a high-resolution RGB image required, or is multispectral data required? What is the ground sampling distance (GSD) requirement for the image? What is the required forward and side overlap rate? The node morphology contains its priority label (such as Critical, High, Medium, Low) in the entire task, which will directly affect the weight of subsequent resource allocation.

[0060] Based on the node shape, corresponding node location, and surrounding rice morphology of the key inspection node, the node influence range of the key inspection node is determined. Based on the "observation height" and camera parameters (field of view, FOV) in the node shape, a basic ground projection rectangle can be calculated. For example, at a height of 8 meters, the UAV's camera FOV covers a ground area of ​​20m x 15m, which is the initial value of the influence range.

[0061] The system analyzes the rice morphology around the node location, including canopy density (if the canopy is dense, the effective range is reduced), ridge orientation and structure (the affected range is designed as a rectangle along the ridge orientation), and stress boundary (if the node is located in a stress area, the affected range will be corrected to an irregular polygon that matches the height of the boundary). At the same time, based on the calculated range, the system adds a dynamic buffer according to the "overlap rate" requirement in the node morphology. For example, in order to achieve 70% lateral overlap, the actual width of the affected range needs to be approximately (1 / (1-0.7)) ≈ 3.33 times the width of the camera's field of view to ensure sufficient overlap between images of adjacent flight paths.

[0062] The system treats the "node influence range" of all key nodes as spatial objects and performs overlay analysis in a unified geographic coordinate system. If the influence ranges of two or more nodes intersect or overlap spatially, the system will merge them into a larger, continuous area. The merging criteria include: node proximity (nodes with similar locations tend to merge their influence ranges), rice distribution consistency (merging is reasonable if the rice distribution characteristics are consistent within the overlapping area), and task synergy (if multiple nodes have similar task types, their ranges can be merged). After merging, the system smooths and optimizes the boundaries of the merged polygon, removing narrow or jagged edges to form the final "key inspection area". Each area is assigned a unique ID and associated with the list of key nodes that make it up.

[0063] Specifically, for the dynamic inspection scenario of rice paddies by drones, the system detected a critical inspection node N_SW_Center located in the center of the stress area in the southwest corner. After analysis, its "node morphology" was determined as follows: the task type is "stem and leaf microscopic detection" because the area shows an abnormally low value on the NDVI map, requiring close-range investigation of the cause; the observation height and attitude are 8 meters, with the camera vertically downward (-90°); the data acquisition specifications require RGB images with GSD better than 1mm / pixel, a forward overlap rate of 80%, and a lateral overlap rate of 70%; the priority is marked as "Critical". These attributes together constitute the "profile" of node N_SW_Center, which was identified by the system as a microscopic detection task point that requires high-precision, low-altitude, and high-overlap imaging.

[0064] For node N_SW_Center, at a height of 8 meters, its basic coverage area is 20m x 15m. However, the system analysis of the initial distribution map of this area revealed that the rice canopy is dense and the NDVI anomaly area is an irregular ellipse with a major axis of approximately 25 meters. Therefore, the system abandoned the 20x15 rectangle and instead adopted a circle with a radius of approximately 15 meters that can completely cover the elliptical stress area as the initial range. To meet the 70% lateral overlap rate, the system calculated that multiple flight paths are needed to cover this circular area. The "node influence range" of this node is defined as a circular area with a diameter of approximately 30 meters, which ensures that when the UAV flies within this area, the images collected are highly correlated with the core area of ​​the anomaly.

[0065] In the farmland, besides the core node N_SW_Center, there are three micro-detection nodes with a priority of "High" in the southwest corner of the stress area. Their "node influence range" (a circle with a diameter of about 20 meters) overlaps to varying degrees with the "node influence range" (a circle with a diameter of 30 meters) of N_SW_Center. The system superimposes these four circular ranges together. Since they are geographically close, all have the task type of "micro-detection", and are all located within the same stressed rice distribution area, the system determines that they should be merged. The system calculates the union of these four overlapping circles, generating an irregular continuous polygon with an area of ​​about 800 square meters. This polygon is defined as the "Key Inspection Area-01" of the farmland. It is associated with the four key inspection nodes and becomes the direct target for the next step of planning the sub-inspection route.

[0066] Furthermore, based on the detection of key inspection areas, the corresponding rice is identified and its current morphology is marked. A first sub-inspection route is determined based on the key inspection area and the morphology of the corresponding rice. A second sub-inspection route is determined based on the key inspection area and the flight pattern of the drone. A sub-inspection route is determined based on the first sub-inspection route, the second sub-inspection route, and the current battery level of the drone. This overall consideration of the first sub-inspection route, the second sub-inspection route, and the current battery level of the drone ensures the accuracy of the sub-inspection routes.

[0067] At this point, the system uses the initial distribution map or real-time preview data to perform a detailed morphological analysis of rice in key inspection areas, including canopy structure (row direction, row spacing, canopy height and density) and stress characteristics (precisely marking the boundaries and morphology of stressed patches). Based on this rice morphological information, the system plans a route that best captures these characteristics. In terms of path mode selection, if the rice rows are obvious, the optimal path is a "parallel row path," that is, the flight path is parallel to the row direction to reduce shadow occlusion. If the stressed patch morphology is irregular, the system adopts an "adaptive grid path," which densifies the flight path within the patch. The spacing of the route is strictly calculated based on the image overlap rate and camera field of view required by the task to ensure no data gaps.

[0068] "Flight morphology" is a quantitative description of the UAV's performance, including kinematic constraints (maximum / minimum speed, minimum turning radius), dynamic constraints (smoothness requirements for attitude changes), and energy consumption models (power consumption curves under different flight states). Based on these constraints, the system plans a smooth path with continuous curvature. For example, the algorithm avoids sharp right-angle turns and instead generates a smooth "cycloid" or "Bezier curve" to connect adjacent routes, enabling the UAV to turn smoothly at a constant angular velocity. At the same time, the system assigns an optimal speed to each point on the path, flying at an economical cruising speed on straight sections and automatically decelerating on turning sections to minimize energy consumption.

[0069] The system weights and fuses the first route ("data-optimal") and the second route ("flight-optimal"), with the weights determined by the mission strategy. The fusion algorithm searches for a new path that satisfies the characteristics of both routes as much as possible. The most crucial step is checking the power constraint: based on the length, speed profile, and energy consumption model of the fused route, the system accurately calculates the "estimated power consumption" required to execute the route. The final decision is as follows:

[0070] Sufficient power: If "current power" > "estimated power consumption" + "safe return power reserve", the fusion route is confirmed as the final "sub-inspection route"; Critical power: If the power is only enough to execute the task but not enough to reserve enough power for safe return, the system will trigger "energy saving mode", reduce the weight of the data route, regenerate a fusion path that is closer to the flight route, or directly reduce the task specifications (such as relaxing the image overlap rate) to reduce energy consumption; Insufficient power: If the current power is not even enough to execute the simplified task, the system will abandon the area, mark it as "pending execution", and replan or return directly.

[0071] Specifically, the system performs detection on "Key Inspection Area-01" (southwest corner stress area); by analyzing high-definition preview images, the system identifies that the rice rows are oriented north-south with a row spacing of approximately 30 centimeters, and accurately marks a "suspected rice blast patch" with a yellowish color and irregular shape in the central area; in order to best observe this patch, the system plans a north-south "bow-shaped" path parallel to the row direction; above the suspected patch, the spacing between the flight paths is increased to 3 meters to ensure extremely high image resolution and overlap; in the healthy area surrounding the patch, the spacing between the flight paths is increased to 5 meters; the entire route is designed entirely around the goal of "how to obtain disease characteristics most efficiently and clearly".

[0072] The system retrieves the drone's "flight profile" data: the minimum turning radius is 5 meters, the economical cruising speed is 3 meters per second, and it needs to be reduced to 1.5 meters per second when turning. Based on the north-south "bow-shaped" path generated by S122.1, the system optimizes each turning point, replacing the original right-angle turns with a circular arc transition with a radius of 5 meters. At the same time, a speed profile is generated: flying at 3 meters per second on a 30-meter straight line, smoothly decelerating to 1.5 meters per second before entering the turn, and then smoothly accelerating after completion. Although the total flight distance may be slightly increased by this route, the flight maneuvers are smoother and the energy consumption is lower.

[0073] The system prepares to generate the final route for "Key Inspection Area-01". It sets the data weight to 70% and the flight weight to 30%, generating a new fusion route: the main body maintains a north-south parallel path, but uses a smooth arc transition at the turns. The system calculates that executing this fusion route will consume 8% of the power. At this time, the drone's "current power" is 45%, and the power reserve for safe return is 20%. Since 45% > 8% + 20%, the condition is met, and this fusion route is finally confirmed as "Sub-inspection Route-01" and issued for execution. If the current power is 28%, the system will determine that the power is insufficient, automatically widen the route spacing, generate a simplified route, and reduce the estimated power consumption to 5%, thereby completing the inspection of the core area while ensuring safety.

[0074] refer to Figure 4 In step S13, the specific steps are as follows:

[0075] S131: The drone acquires the sub-inspection route and conducts a flight inspection along the sub-inspection route. At this time, the drone dynamically takes pictures of the farmland during the flight to collect multiple inspection images. Based on the first recognition of multiple inspection images, the abnormal areas on the surface of the rice are identified. Based on each abnormal area on the surface, the morphology of the rice and the previous pest and disease events of the rice, multiple pest and disease characteristics of the rice are determined.

[0076] S132: Among multiple pest and disease characteristics, the characteristic location and corresponding characteristic morphology of each pest and disease characteristic are determined based on the identification of pest and disease characteristics. At the same time, based on the second identification of multiple inspection images, invasive alien species are identified and the distribution location of invasive alien species in farmland is marked.

[0077] S133: In each key inspection area, the first rice damage coefficient is determined based on the characteristic location of each pest and disease and the distribution location of invasive alien species. The second rice damage coefficient is determined based on the characteristic morphology of each pest and disease and the distribution location of invasive alien species. The rice damage level of the key inspection area is determined based on the mapping relationship between the first rice damage coefficient, the second rice damage coefficient and the rice damage level.

[0078] In the embodiments of this application, the UAV receives the "sub-inspection route" generated by S122 and parses it into a series of navigation instructions with timestamps; the flight controller adjusts the motor speed in real time through advanced control algorithms to ensure that the UAV can accurately track the predetermined route; the camera's shooting is not random, but closely coupled with the flight state, and usually adopts a "distance-triggered" or "time-triggered" mechanism; distance-triggered is more commonly used, that is, whenever the UAV flies forward over a preset distance calculated by the image overlap rate and GSD requirements, a shutter is triggered; for each shutter event, the camera generates an image and immediately writes a set of key metadata into a file, including high-precision spatiotemporal stamps, attitude angles, camera intrinsic and extrinsic parameters, and shooting parameters. This set of data is the basis for all subsequent analyses.

[0079] Before being fed into the recognition model, the image undergoes a series of standardization processes, including color correction, size normalization, and distortion correction based on camera intrinsics and pose angles. The system employs a deep learning model trained on a large number of healthy / abnormal rice images, most commonly U-Net or its variants. This encoder-decoder network, upon receiving the preprocessed image, outputs a "probability map" of the same size as the original image. Each pixel value in the map represents the probability that the pixel belongs to a "surface anomaly." By setting a threshold (e.g., 0.8), this probability can be... Figure 2 Value-encode the data to generate a binary mask; perform connected component analysis on the binary mask to aggregate adjacent anomalous pixels into independent "surface anomalous regions" and calculate their area, centroid, and other geometric properties.

[0080] The system acquires the "morphological" information of rice in the current inspection area, including growth stage (such as tillering stage and heading stage), variety information (resistance information), and canopy structure. Simultaneously, the system queries databases associated with farmland to obtain historical records of past pest and disease events. The system employs an inference engine based on a rule-based or lightweight classification model to match the visual features (such as color, shape, and texture) of "abnormal surface areas" with "rice morphology" and "historical events." For example, a rule might be: IF growth stage = "heading stage" AND abnormal color = "brown spindle-shaped spot" AND historical event contains "rice blast" THEN inferred as "rice blast feature." For each abnormal surface area, if a match is successful, it is upgraded to a "pest and disease feature" and assigned one or more disease type labels and a confidence level.

[0081] Specifically, the drone receives the instruction "Sub-inspection route-01" and takes off from the starting point. It flies along the north-south encrypted route at a speed of 3 meters per second. The flight control system calculates, based on the preset 80% heading overlap rate, that a photo needs to be taken every 4 meters at a flight altitude of 8 meters. Therefore, the drone flies smoothly over the threatened area in the southwest corner of the farmland. Its onboard camera acts like a metronome, clicking every 4 meters to continuously collect dozens of high-resolution RGB images. Each image accurately records the position (e.g., 120.123456° E, 30.654321° N, elevation 25.3 meters) and attitude (pitch -2°, roll 0.5°, heading 180°) at the moment of capture.

[0082] An image captured by a drone is transmitted to the edge computing unit in real time; the U-Net model processes the image and outputs a probability map; in the image, the probability value of the pixels with brown spots on the rice leaves is as high as 0.95, while the probability value of healthy green leaves is less than 0.1; after thresholding and connected component analysis, the system successfully delineates three irregular "surface anomaly regions" on the image and calculates their pixel coordinates and area in the image.

[0083] The system identified three "abnormal surface areas," all of which were brown, spindle-shaped, or irregular in shape. The system determined that the rice in the current farmland was in the "early heading stage," and the variety was "Nanjing 46," which is moderately susceptible to rice blast. The historical database also showed that the farmland had experienced a rice blast outbreak last year. The inference engine was activated, matching the visual feature of "brown spindle-shaped spots" with the contextual information of "heading stage" and "rice blast history." The matching degree was very high, with a confidence level of 92%. These three abnormal surface areas were officially identified as three "disease and pest features" and assigned the following labels: {Feature ID: F001, Type: Rice blast (acute type), Confidence level: 92%}. These features with preliminary diagnostic information will be passed to the next step for more refined analysis.

[0084] Furthermore, among multiple pest and disease characteristics, the feature location and corresponding feature morphology of each pest and disease characteristic are determined based on the identification of the pest and disease characteristics. At the same time, invasive alien species are identified based on the second-level identification of multiple inspection images, and the distribution location of invasive alien species in farmland is marked. This overall consideration of the second-level identification of multiple inspection images is compatible, ensuring the accuracy of invasive alien species identification.

[0085] At this point, the system uses the acquired image metadata (high-precision position, attitude angle) and camera intrinsic parameters to solve the problem through collinearity equations. For each pixel within the pest and disease feature mask, a spatial ray passing through the camera's optical center can be projected backward. When multiple images taken from different locations cover the same feature (this is the purpose of the high overlap design in S122), these rays will intersect at a point in three-dimensional space. Through optimization algorithms such as bundle adjustment, the three-dimensional coordinates (X, Y, Z) of this intersection point, i.e., the "feature position," can be accurately calculated. Simultaneously, the system quantifies the feature morphology, including calculating its geometric metrics (area, perimeter, shape factor, etc.) and optical / texture metrics (statistical features of color and texture) on the image mask, providing data for subsequent severity assessment.

[0086] The system employs a deep learning object detection model, such as YOLO or Faster R-CNN, specifically trained for detecting invasive alien species (e.g., pests, weeds). Unlike segmentation models like U-Net, the object detection model does not concern itself with pixel-level precise contours. Instead, it quickly locates the specific target in the image and encloses it with a bounding box. The model scans each inspection image and outputs one or more detection results, each containing a bounding box, category label, and confidence score. For each detection result with a confidence score higher than a threshold, the system takes the center point of its bounding box and, using the same photogrammetric principle as S132.1, converts its two-dimensional image coordinates into precise three-dimensional geographic coordinates. This coordinate point is the "distribution location" of the invasive alien species.

[0087] Specifically, regarding the "rice blast feature F001" identified in S131, the system found that this feature appeared in three adjacent aerial images. It extracted the shooting position, attitude, and camera parameters of these three images, and used the bundle adjustment algorithm to accurately locate the centroid of F001 to the geographic coordinates (120.123456°E, 30.654321°N, elevation 24.8 meters). This point was marked as the "feature location" of F001. At the same time, the system calculated on the image mask of F001 that its coverage area is approximately 3.14 square centimeters, the shape factor is 0.45 (indicating an irregular shape), the RGB average value is (85,75,60), and it appears as a dark brown color. These quantitative data constitute the "feature morphology" of F001.

[0088] In another image captured by the drone, the YOLO model is performing a second layer of recognition. The model detects a target in the lower right corner of the image, outputs a bounding box, and gives the result: {Category: "Golden apple snail egg mass", Confidence: 0.96, Bounding box: (450, 800, 480, 830)}. Due to the high confidence, the system accepts this detection result. It immediately extracts the center pixel coordinates (465, 815) of the bounding box, and combines them with the image's metadata. Through collinearity equation solving, it obtains the precise geographic coordinates of this golden apple snail egg mass as (120.123459°E, 30.654325°N, elevation 24.5 meters). This point is marked as a "distribution location" and labeled "golden apple snail egg mass".

[0089] Therefore, in each key inspection area, the first rice damage coefficient is determined based on the characteristic location of each pest and disease feature and the distribution location of invasive alien species. The second rice damage coefficient is determined based on the characteristic morphology of each pest and disease feature and the distribution location of invasive alien species. The rice damage level in the key inspection area is determined based on the mapping relationship between the first rice damage coefficient, the second rice damage coefficient, and the rice damage level. This approach takes into account the overall consideration of the mapping relationship between the first rice damage coefficient, the second rice damage coefficient, and the rice damage level, ensuring the accuracy of the rice damage level in the key inspection area. At the same time, multiple key inspection nodes are introduced to further control multiple pest and disease features of rice. This approach takes into account the overall consideration of the characteristic location of each pest and disease feature, the corresponding characteristic morphology, and the distribution location of invasive alien species, improving the accuracy of the rice damage level in the key inspection area.

[0090] At this point, the system aggregates all the "pest and disease characteristic locations" and all the "distribution locations of invasive alien species" determined in S132 to form a unified "damaged location set". Spatial statistical algorithms (most commonly kernel density estimation, KDE) are used to analyze the distribution pattern of this location set. KDE places a kernel function at each location and smoothly superimposes all kernel functions to generate a continuous "damage density map". The system performs statistical analysis on the density map (such as calculating the maximum density value, average density value, or area ratio exceeding the threshold) and normalizes it to a value between 0 and 1, namely the "first rice damage coefficient". The larger the value, the wider and denser the spatial distribution of damage.

[0091] The system assigns different weights to different types of "feature morphology"; for example, the larger the "area" of a lesion, the higher its weight; the darker the "color", the higher its weight; for "invasive alien species", the weight is fixed, but can be adjusted according to their number; for each pest or disease feature in the area, the system calculates an "individual severity score" based on its quantified "feature morphology" and corresponding weight; the scores of all features and the scores of all invasive species are weighted and summed to obtain a "total severity score"; this score is normalized (e.g., mapped to the 0-1 interval) to obtain the final "second rice damage coefficient".

[0092] The rice damage level in the key inspection area is determined based on the mapping relationship between the first rice damage coefficient, the second rice damage coefficient, and the rice damage level. This mapping relationship is the core knowledge base of the system and usually exists in the form of a two-dimensional decision matrix or a classification tree. Its input is the first coefficient (spatial dimension) and the second coefficient (severity dimension), and the output is a discrete "rice damage level". The rule example is as follows: IF first coefficient < 0.3 AND second coefficient < 0.3 THEN level = "Level 1 (healthy)"; IF first coefficient > 0.7 AND second coefficient > 0.7 THEN level = "Level 5 (extremely severe, requiring immediate intervention)"; other combinations correspond to levels 2, 3, and 4, forming a complete decision matrix. The system inputs the two calculated coefficients into this matrix to obtain the final damage level, which is usually accompanied by a brief description and suggestion.

[0093] Specifically, within the "Key Inspection Area-01," the system identified 15 locations with rice blast characteristics and 2 locations with golden apple snail egg masses, totaling 17 damaged points. The system input these 17 geographic coordinates into the KDE algorithm, and the generated density map clearly showed that these points were highly concentrated in the center of the area. The system calculated that the average damage density in this area was 0.82 (after normalization), and the area exceeding the high density threshold accounted for 65% of the total area. Combining these two indicators, the system ultimately calculated the first rice damage coefficient to be 0.85. This high value intuitively indicates that the damage phenomenon is highly concentrated and widespread in space.

[0094] Analysis of 15 rice blast characteristics within the "Key Inspection Area-01" revealed that the lesions of these 15 characteristics had a large average area and were all dark brown in color. According to the preset weighting model, the "individual severity score" of each characteristic was very high, averaging 0.78. Two golden apple snail egg masses each contributed 0.05 to the severity score according to the preset weights. After summing and normalizing all the scores, the system finally calculated the second rice damage coefficient to be 0.80. This high value indicates that even without considering the distribution, each discovered lesion is itself very serious.

[0095] The system has calculated two key coefficients for "Critical Inspection Area-01": the first rice damage coefficient = 0.85 (spatially highly concentrated), and the second rice damage coefficient = 0.80 (high intrinsic severity). The system inputs (0.85, 0.80) into the preset decision matrix. The query rules found that this combination falls into the "Level 5 (Extremely Severe)" range. Therefore, the system finally determines the rice damage level of "Critical Inspection Area-01" to be Level 5 and generates a report: "The rice blast disease in this area is showing an outbreak trend, spatially concentrated and with severe symptoms, rated as extremely severe; it is strongly recommended to take immediate emergency prevention and control measures to prevent further spread." This final diagnosis with precise level and action recommendations will serve as the basis for generating the overall farmland heat map.

[0096] refer to Figure 5 In step S14, the specific steps are as follows:

[0097] S141: Collect the damage level of rice in each key inspection area, mark the location of each key inspection area, and determine the first damage distribution map based on the damage level of rice in each key inspection area and the corresponding regional morphology.

[0098] S142: Determine the second damage distribution map based on the rice damage level of each key inspection area and multiple inspection images, and determine the damage heat map of the key inspection area based on the first and second damage distribution maps.

[0099] S143: Collect the initial distribution map of farmland, determine the optimized distribution map of farmland based on the matching of the initial distribution map of farmland and the regional morphology of key inspection areas, and determine the dynamic distribution map of farmland based on the optimized distribution map of farmland, various damaged heat maps and multiple inspection images.

[0100] In the embodiments of this application, the damage level of rice in each key inspection area is collected, and the location of each key inspection area is marked. The first-level damage distribution map is determined based on the damage level of rice in each key inspection area and the corresponding regional morphology. This approach takes into account the overall consideration of the damage level of rice in each key inspection area and the corresponding regional morphology, ensuring the accuracy of the first-level damage distribution map.

[0101] At this point, the system reads the final conclusions of all completed assessments of the "key inspection areas" in batches from the output of S133, including the unique ID of each area and its corresponding "rice damage level"; the system obtains the "regional morphology" of each key inspection area from the output of S121, that is, its precise geographical boundary (usually represented by one or more closed polygons); the system associates the collected level and location information to form a structured geospatial vector dataset; each record in the dataset contains two fields: one stores geometric information (polygon) and the other stores attribute information (damage level).

[0102] The system needs to establish a "level-color" mapping relationship, which is a preset color scheme designed to convey the severity of damage by utilizing human intuitive perception of color. For example, a standard scheme is: level 1 (healthy) corresponds to dark green, level 2 (mild) corresponds to light green, level 3 (moderate) corresponds to yellow, level 4 (severe) corresponds to orange, and level 5 (extremely severe) corresponds to red. The system iterates through each feature in the vector dataset generated in S141.1, reads its damage level attribute value, and assigns the corresponding fill color to the polygon feature according to the rendering scheme. The system overlays all the symbolized polygon features with a base map (such as a satellite image) to generate a raster image or interactive map, which is the "first-level damage distribution map".

[0103] Specifically, step S133 completed the assessment of all key inspection areas of the farmland; the system collected data and obtained the following results: Area-01 (southwest corner area) damage level 5; Area-02 (central area) damage level 2; Area-03 (northeast corner area) damage level 4; Area-04 (northwest corner area) damage level 1; Area-05 (southeast corner area) damage level 1; at the same time, the system retrieved the precise polygon boundaries of these 5 areas generated in S121 and defined by GPS coordinates; the system created a vector dataset containing 5 polygon features, each feature's attribute table has a field named Damage_Level, and the corresponding level value is filled in.

[0104] The system begins rendering the created vector dataset; it reads Area-01's Damage_Level as 5 and fills the interior of its polygon boundaries with red; it reads Area-03's level as 4 and fills it with orange; it reads Area-02's level as 2 and fills it with light green; Area-04 and Area-05 have levels of 1 and are filled with dark green; the system overlays these five colored polygons onto the satellite base map of the farmland; the resulting "first-level damage distribution map" clearly shows a picture: the southwest corner of the farmland is a striking red, the northeast corner is a warning orange, the central area is a slight light green, and the rest is a healthy dark green. This map provides a clear macro-level understanding of the overall health of the farmland.

[0105] Furthermore, a second damage distribution map is determined based on the damage level of rice in each key inspection area and multiple inspection images. A damage heat map of the key inspection area is then determined based on the first and second damage distribution maps, taking into account both the first and second damage distribution maps to ensure the accuracy of the damage heat map of the key inspection area.

[0106] At this point, the system uses all the "inspection images" collected in S131 and their accompanying high-precision position and attitude data to perform photogrammetric processing; through aerial triangulation and bundle adjustment, it calculates the precise exterior orientation elements of all photos, performs orthorectification on each image, and eliminates geometric distortion caused by camera tilt and terrain undulation; it seamlessly stitches thousands of orthorectified images into an orthophoto map covering the entire farmland with a unified geographic coordinate system; then, the system spatially overlays the vector polygons (regional morphology) and their attributes (damage level) generated in S141 with the orthophoto map, and performs a polygon-to-grid assignment operation: it traverses each pixel of the orthophoto map, determines which region the pixel falls into, and directly assigns the "rice damage level" value of that region to the pixel.

[0107] The system selects a suitable spatial interpolation algorithm, commonly including Inverse Distance Weighting (IDW) and Kriging. Kriging is a more advanced geostatistical method that considers not only distance but also the spatial autocorrelation of known data points, thus providing the optimal linear unbiased estimate, which is usually the smoothest and most geographically accurate. When performing interpolation, the system treats each pixel in the "second-level damage distribution map" as a known sampling point with a "damage level" value. The interpolation algorithm generates a regular grid across the entire farmland based on these sampling points, and estimates a continuous, smooth damage value for each node of the grid based on the distance and spatial correlation of surrounding sampling points. The system again uses the "level-color" rendering scheme, but this time it is applied to a continuous numerical range and color band interpolation is performed, so that pixels with values ​​between 4.0 and 5.0 appear as a gradient between orange and red.

[0108] Specifically, the system processes 800 high-resolution images taken by the drone over the farmland using professional photogrammetry software to generate an orthophoto map with a resolution of 0.5 cm. Then, the system overlays the five colored polygons (red in the southwest corner, orange in the northeast corner, etc.) defined in S141 onto this map. The system performs an assignment operation: on the orthophoto map, all pixels falling within the red polygon in the southwest corner are uniformly set to a value of 5; pixels falling within the orange polygon in the northeast corner are uniformly set to a value of 4, and so on. The system obtains a "second layer of damage distribution map".

[0109] The system uses the Kriging interpolation algorithm to process the "second layer of damage distribution map." The algorithm uses pixels with a value of 5 in the southwest corner and pixels with a value of 2 in the central area as the main input points to analyze their spatial variation trends. At the boundary between the two areas, the algorithm does not produce a harsh boundary line, but calculates a smooth transition zone. The pixel value adjacent to the red area is estimated as 4.8, and further away as 4.5, 4.2, etc., and the color also smoothly transitions from red to dark orange, orange, and light orange, until it seamlessly connects with the light green of the central area. On the final generated "damage heat map," the red area in the southwest corner acts like a "heat source," and its heat (the degree of damage) gradually decreases towards the surrounding areas. This map not only tells managers where the problem is, but also intuitively shows how the problem is spreading to the surrounding areas, greatly enhancing the information value for decision-making.

[0110] Therefore, an initial distribution map of farmland is collected, and an optimized distribution map of farmland is determined based on the matching of the initial distribution map of farmland with the regional morphology of key inspection areas. A dynamic distribution map of farmland is determined based on the optimized distribution map of farmland, various damage heat maps and multiple inspection images. This comprehensive approach takes into account the optimized distribution map of farmland, various damage heat maps and multiple inspection images, ensuring the accuracy of the dynamic distribution map of farmland.

[0111] At this point, the system acquires the "initial distribution map of farmland," which is usually a static GIS data set containing plot boundaries, digital elevation models, and fixed features. The system then performs spatial overlay analysis between the "regional morphology of key inspection areas" (i.e., the boundary polygons determined in S121) generated during this inspection and the "initial distribution map." This matching process is not just a simple alignment but also a data verification and update. The system can correct the initial boundaries by comparing them with high-resolution orthophotos, or use image recognition algorithms to find and add features that are not present in the initial distribution map from the orthophotos. After matching, correction, and updating, the system generates a new, more accurate, and more realistic basic geographic map of farmland, namely the "optimized distribution map."

[0112] The system, within a GIS rendering engine, overlays multiple layers in a logical order to form a layered structure: Bottom layer: The generated "optimized distribution map of farmland," providing a precise geographic framework and background reference; Middle layer (semi-transparent): "Various damage heatmaps" generated by S142. This layer is typically set to semi-transparent so that it clearly displays the distribution of damage without completely obscuring the details of the base map; Top layer (optional / on-demand): A high-resolution orthophoto map generated by stitching together "multiple inspection images." This layer is hidden by default, but users can switch its display at any time to zoom in on specific areas for visual inspection of "ground reality"; Annotation layer: Precisely located, high-confidence "pest and disease characteristic locations" or "distribution locations of invasive alien species" from S132, overlaid on the map with specific icons and clickable attribute information; The final "dynamic distribution map" is not a static image but an interactive GIS product, supporting functions such as zooming and panning, layer switching, information querying, and measurement and analysis.

[0113] Specifically, the system retrieved the initial distribution map of the farmland, which included plot boundaries and a main irrigation canal running north-south. On the orthophoto generated during this inspection, the system found that due to recent field work, the western field ridge had widened by about 1.5 meters, and a temporary pumping point had been added in the southeast corner. The system automatically identified these changes, updated the polygon coordinates of the western field ridge, and added a point feature to the optimized distribution map to represent the new pumping point. The final "optimized distribution map" is an accurate base map reflecting the latest status of the farmland.

[0114] The system ultimately generates a "dynamic distribution map" of the farmland. On this map, users see precise field ridges and irrigation ditches (optimized distribution map); a semi-transparent heat map in red, yellow, and green stripes covers it, with a striking red "heat source" in the southwest corner; when users zoom in on the map to the southwest corner, they can switch to the underlying orthophoto map, clearly seeing the details of each rice plant; at the same time, 15 red rice blast icons and 2 purple golden apple snail egg mass icons are precisely marked on the image; clicking on one of the red icons will bring up an information box: "Feature ID: F001, Type: Rice Blast (Acute), Confidence: 92%, Area: 3.14cm²". This final "dynamic distribution map" is exported in a standard format and uploaded to the farmland management cloud platform; farm administrators can intuitively grasp the overall situation through a mobile app or computer, and accurately locate each problem point, thereby formulating the most precise variable spraying or manual intervention plan.

[0115] refer to Figure 6 In step S15, the specific steps are as follows:

[0116] S151: Collect dynamic distribution map, determine multiple abnormal rice nodes based on the detection of the dynamic distribution map, and determine multiple abnormal rice regions based on the relative position of each abnormal rice node, the corresponding node shape and the distribution of the dynamic distribution map. At this time, each abnormal rice region is distributed in a different position on the dynamic distribution map.

[0117] S152: Based on the detection of each abnormal rice area, determine the water flow area associated with the abnormal rice area; determine the first abnormal inspection content based on the regional distribution of the water flow area and the regional location of multiple abnormal rice areas; determine the second abnormal inspection content based on the regional location of multiple abnormal rice areas and the current location of the drone; and determine the abnormal inspection event of the drone based on the first and second abnormal inspection contents.

[0118] S153: Collect multiple weather and environmental data of farmland, determine the weather conditions corresponding to the farmland based on the multiple weather and environmental data and the dynamic distribution map of the farmland, determine multiple key inspection events based on the weather conditions corresponding to the farmland and the abnormal inspection events of the drone, and determine the dynamic inspection events of the drone based on the multiple key inspection events, the regional location of each abnormal rice area and the corresponding regional morphology.

[0119] In the embodiments of this application, a dynamic distribution map is collected, and multiple abnormal rice nodes are determined based on the detection of the dynamic distribution map. Multiple abnormal rice regions are determined based on the relative position of each abnormal rice node, the corresponding node shape, and the distribution of the dynamic distribution map. At this time, each abnormal rice region is distributed in a different position on the dynamic distribution map, which takes into account the overall consideration of the relative position of each abnormal rice node, the corresponding node shape, and the distribution of the dynamic distribution map, and ensures the accuracy of multiple abnormal rice regions.

[0120] At this point, the system acquires the "damaged heatmap" in raster format generated by S142, where each pixel value represents a continuous damage assessment value; detection is performed using threshold segmentation technology; the system sets one or more anomaly judgment thresholds (e.g., 4.0, representing "severe damage" and above), and iterates through each pixel of the heatmap, comparing its pixel value with the threshold; all pixels with pixel values ​​greater than or equal to the threshold are marked as "abnormal rice nodes"; in technical implementation, this usually involves generating a binary mask image the same size as the original image, where pixels with a value of 1 correspond to abnormal nodes, and pixels with a value of 0 correspond to normal areas. These nodes are the most basic units constituting abnormal areas.

[0121] The system performs spatial clustering (connectivity analysis) on the generated binary mask image. The algorithm finds all interconnected anomalous nodes and groups them into the same cluster. Each independent cluster constitutes a preliminary "abnormal rice region". Next, the system applies morphological processing (such as "closing operation" to fill small holes and "opening operation" to eliminate noise points) to optimize the region boundary. For each processed cluster, the system calculates its region morphology, including centroid, area, bounding box, and shape descriptor. The system also analyzes the "distribution" of these newly generated anomalous regions on the entire "dynamic distribution map" to determine whether they are isolated point distributions, linear distributions along a line, or area distributions.

[0122] Furthermore, based on the detection of each abnormal rice area, the associated water flow area is determined. The first abnormal patrol content is determined according to the regional distribution of the water flow area and the regional location of multiple abnormal rice areas. The second abnormal patrol content is determined according to the regional location of multiple abnormal rice areas and the current location of the drone. The abnormal patrol event of the drone is determined based on the first and second abnormal patrol content. This method takes into account both the first and second abnormal patrol content as a whole, ensuring the accuracy of the abnormal patrol event of the drone.

[0123] At this point, the system extracts the vector layer of hydrological elements (such as irrigation canals and drainage ditches) from the "optimized distribution map" generated in S143; it performs spatial proximity analysis on the centroid or boundary of each "abnormal rice area" with the hydrological network layer; the most commonly used technique is buffer analysis: the system creates a buffer with a specified radius centered on each anomalous area and queries which hydrological elements intersect with this buffer; if multiple anomalous areas are associated with the same or interconnected hydrological elements, the system infers that they share a "related flow area"; furthermore, the system can use network analysis, combined with digital elevation model (DEM) data, to determine the upstream and downstream relationships of these anomalous areas on the flow path.

[0124] Based on the regional distribution of the water flow area and the location of multiple abnormal rice areas, the first anomaly inspection content is determined. The core of the first anomaly inspection content (based on insight) is "source tracing". If the associated water flow area is identified, the system will generate a task to trace upstream along the water flow path. The inspection task is defined as "high-precision reconnaissance of the upstream section of the associated water flow area". Specific tasks include reducing the flight altitude, taking pictures of the water surface, or taking detailed images of the rice on both sides of the upstream channel.

[0125] Based on the regional locations of multiple abnormal rice areas and the current location of the drone, the second anomaly inspection content is determined. The core of the second anomaly inspection content (based on efficiency) is "path optimization". It is entirely based on geometric calculation and does not consider agronomic logic. The system obtains the centroid coordinates of all abnormal areas and the "current location" transmitted back by the drone in real time, and runs a heuristic algorithm for solving the Traveling Salesman Problem (TSP) to find the shortest or least energy-consuming path from the current location to visit all abnormal areas.

[0126] The system logically integrates the "first anomaly inspection content" (source tracing task) and the "second anomaly inspection content" (efficiency path). Typically, the source tracing task has higher agronomical value and is therefore given higher priority. The integration strategy is to use the efficiency path as the "skeleton" and insert the source tracing task as the "core module" into the most appropriate position in the path. The system generates a structured "anomaly inspection event" object, which contains complete information such as event ID, priority, task description, detailed waypoint list (including coordinates, altitude, and actions), and success conditions.

[0127] Therefore, multiple weather and environmental data points for farmland are collected. Based on these data points and a dynamic distribution map of the farmland, the corresponding weather conditions are determined. Based on the weather conditions and abnormal drone patrol events, multiple key patrol events are identified. Based on these key patrol events, the location of each abnormal rice paddy area, and its corresponding morphology, dynamic drone patrol events are determined. This approach considers the overall situation of multiple key patrol events, the location of each abnormal rice paddy area, and its corresponding morphology, ensuring the accuracy of the dynamic drone patrol events. Furthermore, the introduction of a dynamic farmland distribution map further manages abnormal drone patrol events, achieving a holistic consideration of abnormal drone patrol events, each abnormal rice paddy area, and the corresponding weather conditions for the farmland, thus improving the accuracy of the dynamic drone patrol events.

[0128] At this time, the system collects weather data (probability of future rainfall, humidity, temperature, etc.) and environmental data (soil moisture, leaf surface moisture, etc.) in real time through API interfaces or sensor networks; it integrates and analyzes this data with the "dynamic distribution map of farmland"; the system runs a multi-factor risk assessment model based on agronomic knowledge or machine learning. The input of this model includes disease type, current damage level, air humidity, probability of future rainfall, etc.; the model outputs a comprehensive and descriptive "weather conditions corresponding to farmland" according to preset rules (e.g.: IF disease type="rice blast" AND air humidity > 90% AND probability of rainfall in the next 6 hours > 80% THEN risk level="extremely high").

[0129] The system prioritizes "abnormal inspection events" based on the risk level of "weather conditions." An event that was originally prioritized as "medium" will be upgraded to "urgent" if it encounters "extremely high risk" weather. At the same time, high-risk weather will generate new and more specific inspection tasks. The system will add new sub-tasks to the original events based on the risk type. For example, if the risk is the spread of disease, the system will add a task to "check whether the drainage outlets of all abnormal areas and related water flow areas are unobstructed." After priority adjustment and content enhancement, the system generates one or more "key inspection events" with clear priority labels and detailed task lists.

[0130] If multiple "key inspection events" exist, the system will plan the task sequence and determine the execution order based on their priority, geographical location, and drone battery level. For each key inspection event to be executed, the system will perform detailed flight path planning based on the "regional location of the abnormal rice paddy area" and "regional morphology." For example, for large, irregularly shaped abnormal areas, the system will plan a "bow-shaped" or "spiral" coverage flight path; for small but critical areas (such as drainage outlets), a fixed-point action of "hovering-zooming-taking photos" will be planned. The system uses a path planning algorithm to generate the safest and most energy-efficient three-dimensional flight path connecting all task points, taking into account obstacles and no-fly zones, and packages it into a complete instruction set that can be directly issued to the drone flight control system, which is the final "dynamic inspection event."

[0131] Please see Figure 7 , Figure 7 This is a schematic diagram of the structural composition of the drone-based dynamic inspection device for farmland according to an embodiment of the present invention; the drone-based dynamic inspection device for farmland includes:

[0132] The key inspection node module 21 is used to determine the inspection path of the UAV on the farmland based on the initial distribution map of the farmland and the current position of the UAV, and to determine multiple key inspection nodes based on the inspection path and the distribution pattern of rice; the multiple key inspection nodes are located at different heights.

[0133] The sub-inspection route module 22 is used to determine the key inspection area based on the node position and the corresponding node influence range of each key inspection node, and to determine the sub-inspection route based on the key inspection area, the corresponding rice and drones.

[0134] The rice damage level module 23 is used for the UAV to fly along the sub-inspection route and collect multiple inspection images. Based on the multiple inspection images and the morphology of the rice, it determines multiple pest and disease characteristics of the rice. Based on the characteristic location of each pest and disease characteristic, the corresponding characteristic morphology, and the distribution location of invasive alien species, it determines the rice damage level of the key inspection area.

[0135] The dynamic distribution map module 24 is used to determine the damage heat map of each key inspection area based on the rice damage level, corresponding regional location and multiple inspection images, and to determine the dynamic distribution map of the farmland based on each damage heat map, the initial distribution map of the farmland and the regional morphology of the key inspection area.

[0136] The dynamic patrol event module 25 is used to identify multiple abnormal rice areas based on the identification of dynamic distribution maps, determine the abnormal patrol events of the drone based on the regional location of multiple abnormal rice areas, the water flow areas associated with the abnormal rice areas and the current location of the drone, and determine the dynamic patrol events of the drone based on the abnormal patrol events of the drone, the weather conditions corresponding to each abnormal rice area and farmland.

[0137] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

Claims

1. A method for dynamic inspection of farmland using unmanned aerial vehicles (UAVs), characterized in that, include: Based on the initial distribution map of the farmland and the current location of the drone, the inspection path of the drone for the farmland is determined, and based on the inspection path and the distribution pattern of rice, several key inspection nodes are determined. Multiple key inspection nodes are located at different heights; The key inspection area is determined based on the node location and the corresponding node influence range of each key inspection node, and the sub-inspection route is determined based on the key inspection area, the corresponding rice and drones. The drone flies along the sub-inspection route and collects multiple inspection images. Based on the multiple inspection images and the morphology of the rice, it determines multiple pest and disease characteristics of the rice. Based on the characteristic location of each pest and disease characteristic, the corresponding characteristic morphology, and the distribution location of invasive alien species, it determines the damage level of the rice in the key inspection area. Based on the rice damage level, corresponding location, and multiple inspection images of each key inspection area, a damage heat map of that key inspection area is determined. Based on these damage heat maps, the initial farmland distribution map, and the regional morphology of the key inspection areas, a dynamic farmland distribution map is determined. This process includes: collecting the rice damage level of each key inspection area and marking the location of each key inspection area; determining a first-level damage distribution map based on the rice damage level and corresponding regional morphology of each key inspection area; determining a second-level damage distribution map based on the rice damage level of each key inspection area and multiple inspection images; determining a damage heat map of the key inspection area based on the first and second-level damage distribution maps; collecting the initial farmland distribution map; determining an optimized farmland distribution map based on the matching of the initial farmland distribution map and the regional morphology of the key inspection areas; and determining a dynamic farmland distribution map based on the optimized farmland distribution map, the damage heat maps, and multiple inspection images. Multiple layers are overlaid in a logical order to form a layered structure: Bottom layer: the generated "optimized distribution map of farmland"; Middle layer: the generated "various damage heatmaps," this layer is usually set to semi-transparent so that it clearly shows the distribution of damage without completely obscuring the details of the base map; Top layer: a high-resolution orthophoto map generated by stitching together "multiple inspection images," this layer is hidden by default, but users can switch its display at any time to zoom in on specific areas; Annotation layer: located, high-confidence "pest and disease characteristic locations" or "distribution locations of invasive alien species," with specific... Defined icons are overlaid on the map, along with clickable attribute information. The final "dynamic distribution map" is not a static image but an interactive GIS product, supporting zooming and panning, layer switching, information querying, and measurement and analysis functions. Based on the identification of the dynamic distribution map, multiple abnormal rice areas are determined. Based on the regional locations of these abnormal rice areas, the associated water flow areas, and the current location of the drone, abnormal drone patrol events are determined. Based on the abnormal drone patrol events, the weather conditions corresponding to each abnormal rice area and farmland, dynamic drone patrol events are determined.

2. The method for dynamic inspection of farmland by unmanned aerial vehicles according to claim 1, characterized in that, The inspection path of the UAV to the farmland is determined based on the initial distribution map of the farmland and the current location of the UAV, and multiple key inspection nodes are determined based on the inspection path and the distribution pattern of rice. Multiple key inspection nodes are located at different heights, including: Locate the farmland and collect an initial distribution map of the farmland; determine the drone's flight area based on the farmland's location, the corresponding initial distribution map, and the drone's current location; and determine the drone's inspection path for the farmland based on the drone's flight area and the distribution pattern of the farmland. Based on the detection of the inspection path, multiple sub-inspection paths are determined. At the same time, the distribution style of rice in the farmland is collected. Multiple inspection nodes are determined according to the multiple sub-inspection paths, the position difference between two adjacent sub-inspection paths, and the distribution style of rice. Multiple key inspection nodes are determined based on the node positions and corresponding node priorities of multiple inspection nodes. These key inspection nodes are located in different positions in the farmland, covering both the internal and upper spaces of the farmland, and are situated at different heights.

3. The method for dynamic inspection of farmland by unmanned aerial vehicles according to claim 1, characterized in that, The process involves determining key inspection areas based on the node locations and corresponding influence ranges of each key inspection node, and then determining sub-inspection routes based on these key inspection areas, the corresponding rice paddies, and drones. This includes: Among multiple key inspection nodes, the node shape of each key inspection node is determined based on the detection of each key inspection node. The node influence range of the key inspection node is determined based on the node shape, the corresponding node position, and the shape of the surrounding rice. The key inspection area is determined according to the node position, the corresponding node influence range, and the distribution range of the surrounding rice of each key inspection node. Based on the detection of key inspection areas, the corresponding rice is identified and its current form is marked. A first sub-inspection route is determined based on the key inspection area and the form of the corresponding rice. A second sub-inspection route is determined based on the key inspection area and the flight pattern of the drone. A sub-inspection route is determined based on the first sub-inspection route, the second sub-inspection route, and the current battery level of the drone.

4. The method for dynamic inspection of farmland by unmanned aerial vehicles according to claim 1, characterized in that, The drone flies along the sub-inspection route and collects multiple inspection images. Based on these images and the morphology of the rice, it determines multiple pest and disease characteristics of the rice. Based on the characteristic location of each pest and disease characteristic, its corresponding morphological features, and the distribution location of invasive alien species, it determines the damage level of the rice in the key inspection area, including: The drone acquires the sub-inspection route and conducts a flight inspection along the sub-inspection route. During the flight, the drone dynamically photographs the farmland to collect multiple inspection images. Based on the first identification of multiple inspection images, abnormal areas on the surface of the rice are identified. Based on each abnormal surface area, the morphology of the rice, and previous pest and disease events of the rice, multiple pest and disease characteristics of the rice are determined. Among multiple pest and disease characteristics, the characteristic location and corresponding characteristic morphology of each pest and disease characteristic are determined based on the identification of pest and disease characteristics. At the same time, based on the second identification of multiple inspection images, invasive alien species are identified and the distribution location of invasive alien species in farmland is marked.

5. The method for dynamic inspection of farmland by unmanned aerial vehicle (UAV) according to claim 4, characterized in that, The drone flies along a sub-inspection route and collects multiple inspection images. Based on these images and the morphology of the rice, it determines multiple pest and disease characteristics of the rice. Based on the characteristic location of each pest and disease characteristic, its corresponding morphological features, and the distribution location of invasive alien species, it determines the damage level of the rice in the key inspection area. The method also includes: In each key inspection area, the first rice damage coefficient is determined based on the characteristic location of each pest and disease and the distribution location of invasive alien species. The second rice damage coefficient is determined based on the characteristic morphology of each pest and disease and the distribution location of invasive alien species. The rice damage level of the key inspection area is determined based on the mapping relationship between the first rice damage coefficient, the second rice damage coefficient and the rice damage level.

6. The method for dynamic inspection of farmland by unmanned aerial vehicles according to claim 1, characterized in that, The identification of multiple abnormal rice areas based on dynamic distribution maps, the determination of abnormal patrol events by the drone based on the regional locations of these abnormal rice areas, the associated water flow areas, and the current location of the drone, and the determination of dynamic patrol events by the drone based on these abnormal patrol events, the weather conditions corresponding to each abnormal rice area and farmland, including: A dynamic distribution map is collected, and multiple abnormal rice nodes are identified based on the detection of the dynamic distribution map. Multiple abnormal rice regions are identified based on the relative position of each abnormal rice node, the corresponding node morphology, and the distribution of the dynamic distribution map. At this time, each abnormal rice region is distributed in a different position on the dynamic distribution map. Based on the detection of each abnormal rice area, the associated water flow area is determined. Based on the regional distribution of the water flow area and the regional location of multiple abnormal rice areas, the first abnormal patrol content is determined. Based on the regional location of multiple abnormal rice areas and the current location of the drone, the second abnormal patrol content is determined. Based on the first and second abnormal patrol content, the abnormal patrol event of the drone is determined.

7. The method for dynamic inspection of farmland by unmanned aerial vehicle (UAV) according to claim 6, characterized in that, The method of identifying multiple abnormal rice-growing areas based on dynamic distribution maps, determining abnormal drone patrol events based on the regional locations of these abnormal rice-growing areas, the associated water flow areas, and the current location of the drone, and determining dynamic drone patrol events based on the abnormal drone patrol events, the weather conditions corresponding to each abnormal rice-growing area and farmland, also includes: Multiple weather and environmental data points for farmland are collected. Based on these data points and the dynamic distribution map of the farmland, the corresponding weather conditions are determined. Based on the weather conditions for the farmland and abnormal patrol events by drones, multiple key patrol events are identified. Based on these key patrol events, the location of each abnormal rice area, and the corresponding regional morphology, dynamic patrol events by drones are determined.

8. A dynamic inspection device for farmland using unmanned aerial vehicles (UAVs), characterized in that, The drone-based dynamic inspection device for farmland is applied to the drone-based dynamic inspection method for farmland as described in any one of claims 1-7, wherein the drone-based dynamic inspection device for farmland comprises: The key inspection node module is used to determine the inspection path of the UAV on the farmland based on the initial distribution map of the farmland and the current position of the UAV. Based on the inspection path and the distribution pattern of rice, multiple key inspection nodes are determined; the multiple key inspection nodes are located at different altitudes. The sub-inspection route module is used to determine the key inspection area based on the node location and the corresponding node influence range of each key inspection node, and to determine the sub-inspection route based on the key inspection area, the corresponding rice and drones. The rice damage level module is used by drones to fly along sub-inspection routes and collect multiple inspection images. Based on the multiple inspection images and the morphology of rice, multiple pest and disease characteristics of rice are determined. Based on the characteristic location of each pest and disease characteristic, the corresponding characteristic morphology, and the distribution location of invasive alien species, the damage level of rice in the key inspection area is determined. The dynamic distribution map module is used to determine the damage heat map of each key inspection area based on the rice damage level, corresponding regional location, and multiple inspection images. Based on each damage heat map, the initial distribution map of farmland, and the regional morphology of the key inspection area, the dynamic distribution map of farmland is determined. The dynamic patrol event module is used to identify multiple abnormal rice areas based on the dynamic distribution map. It determines the abnormal patrol events of the drone based on the regional location of multiple abnormal rice areas, the water flow areas associated with the abnormal rice areas, and the current location of the drone. It determines the dynamic patrol events of the drone based on the abnormal patrol events of the drone, the weather conditions corresponding to each abnormal rice area and farmland.

Images