An unmanned aerial vehicle-based vegetation AI identification management method and system

Abstract

The embodiment of the present application relates to the technical field of plant management, and discloses a vegetation AI identification management method based on a UAV, comprising: identifying wide-area remote sensing data to obtain all abnormal sub-regions and window detection regions in a target detection region; performing a second collection task on all abnormal sub-regions and window detection regions to obtain corresponding fine remote sensing data; determining vegetation state information of the abnormal sub-regions according to the fine remote sensing data, and determining a canopy shielding relationship in the window detection regions according to the fine remote sensing data, and identifying and determining a potential under-forest region under the canopy coverage; performing probabilistic inference on the vegetation state of the potential under-forest region by fusing multiple sources of evidence, generating an under-forest vegetation probability distribution map containing an inferred vegetation category and a corresponding confidence; and generating a comprehensive vegetation state report of the target detection region. The overall identification management efficiency is improved by means of hierarchical collection.

Description

Technical Field

[0001] This invention relates to the field of plant management technology, specifically to a method and system for vegetation AI recognition and management based on unmanned aerial vehicles (UAVs). Background Technology

[0002] Currently, vegetation and forestry status identification mainly relies on manual inspections or satellite remote sensing, but these methods suffer from drawbacks such as low efficiency, insufficient identification accuracy, and inability to provide timely data feedback. Therefore, designing an efficient vegetation identification and management solution has become a pressing technical problem for those skilled in the art. Summary of the Invention

[0003] To address the aforementioned shortcomings, this invention discloses a vegetation AI recognition and management method based on unmanned aerial vehicles (UAVs), which can achieve efficient vegetation recognition and management.

[0004] The first aspect of this invention discloses a method for vegetation AI recognition and management based on unmanned aerial vehicles (UAVs), comprising: The first UAV is controlled to perform the first data acquisition task at the first flight altitude to acquire wide-area remote sensing data of the target detection area; and the wide-area remote sensing data is identified to obtain all abnormal sub-regions and window detection areas within the target detection area; The second acquisition task is determined based on the spatial location and area information of all abnormal sub-regions and window detection areas. Control the second UAV to perform a second acquisition task at a second flight altitude lower than the first flight altitude, and perform a second acquisition task on all abnormal sub-regions and window detection areas to obtain corresponding fine remote sensing data; Based on the fine remote sensing data, vegetation status information of abnormal sub-regions is determined, and canopy shading relationships in the window detection area are determined based on the fine remote sensing data, and potential understory areas under canopy cover are identified. By fusing multi-source evidence, the vegetation status of potential understory areas is probabilistically inferred, generating a probability distribution map of understory vegetation that includes inferred vegetation categories and corresponding confidence levels; wherein, the multi-source evidence includes direct observation data of window detection areas identified from the fine remote sensing data, partial exposure features of canopy edge areas, and association rules based on ecological prior knowledge; Based on the wide-area status information of the wide-area remote sensing data, the refined identification results of the fine remote sensing data, and the inferred information of the forest understory vegetation probability distribution map, a comprehensive vegetation status report of the target detection area is generated.

[0005] As an optional implementation, in a first aspect of the present invention, determining the canopy shading relationship in the window detection area based on the fine remote sensing data, and identifying potential understory areas under the canopy cover, includes: Semantic segmentation is performed on the refined remote sensing data to identify and segment canopy gaps; spectral and texture features are extracted from each window detection area, and the vegetation in each window detection area is classified using a first vegetation classification model to obtain a first set of observed species and their spatial locations. Identify the outline edges of each canopy object in the refined remote sensing data, and extract the edge transition zone of a preset width; analyze the image features within the edge transition zone, and identify the partially exposed understory vegetation features through a second vegetation classification model to obtain a second set of observed species; A spatial probability propagation model is established, using the species information in the first and second observed species sets as known evidence sources. Based on the distance decay function and the canopy density decay coefficient, the initial probability of each species existing at each location in the potential understory area is calculated. An ecological knowledge graph is invoked, which includes rules on species symbiosis, light requirements, and terrain preferences. Based on the ecological knowledge graph, the initial probabilities are constrained and adjusted to generate the probability distribution map of the understory vegetation.

[0006] As an optional implementation, in a first aspect of the present invention, identifying the wide-area remote sensing data to obtain all window detection regions within the target detection area includes: The wide-area remote sensing data is preprocessed to generate a canopy cover binarized mask image; The seed connected region growing algorithm is used to mark and segment the non-canopy region in the canopy cover binarized mask image to identify independent window patches; Calculate the area, perimeter, minimum bounding rectangle, and shape complexity index of each window patch; Alternatively, the wide-area remote sensing data can be identified to obtain all window detection regions within the target detection area, including: Identify all window detection regions within the target area and calculate the window area and geometric features of each window detection region; Perform a second acquisition task on all window detection areas to obtain the corresponding detailed remote sensing data, including: Based on the window area of ​​the target detection region and the geometric features, the corresponding second UAV is dynamically determined and assigned to perform the second data acquisition task at the second flight altitude. For window detection areas with a window area larger than a preset window threshold, a single UAV performs a detailed grid survey. For multiple window detection areas with a window area smaller than the preset window threshold and discretely distributed, the corresponding acquisition points are determined based on the positions between each window detection area so that the second UAV can acquire multiple window detection areas at the second flight altitude; and the corresponding second acquisition task is performed to acquire the corresponding fine remote sensing data.

[0007] As an optional implementation, in a first aspect of the present invention, determining the canopy shading relationship in the window detection area based on the fine remote sensing data, and identifying potential understory areas under the canopy cover, includes: The detailed remote sensing data is analyzed to determine the causal type of windows in the window detection area, including a first type of window formed by gaps between adjacent trees and a second type of window formed by openings in the canopy of a single tree. Based on the aforementioned cause types, different spatial reasoning models are adopted. For the first type of window, it is treated as an independent observation window, and its internal vegetation is directly classified. For the second type of window, it is associated with the canopy model of the tree to which it belongs, and its edge features are analyzed to infer the overall canopy health status and lower light transmission conditions of the tree. By integrating direct classification and inference results, a forest understory vegetation status map containing vegetation categories and corresponding confidence levels is generated.

[0008] As an optional implementation, in a first aspect of the present invention, before controlling the first UAV to perform the first data collection task at the first flight altitude, the method further includes: Obtain the boundary range of the target area and multiple pre-divided or automatically generated management plot units within it; Based on preliminary analysis of historical data, land cover maps or initial remote sensing data, each management plot unit is assigned a corresponding attribute label, which includes one or more of the following: vegetation type, functional zoning, management priority, estimated window density, and historical pest and disease incidence rate. Based on the attribute tags, an initial set of drone operation parameter suggestions is generated for each of the managed plot units. The set of operation parameter suggestions includes recommended flight altitude, sensor type, flight path overlap rate, and whether to enable cluster collaboration mode.

[0009] As an optional implementation, in a first aspect of the present invention, identifying the wide-area remote sensing data to obtain all abnormal sub-regions within the target detection area includes: The wide-area remote sensing data is identified to obtain abnormal pixels or abnormal patches within the target detection area. Based on their geographic coordinates, a density-based spatial clustering algorithm is used to group them, and abnormal points with a spatial distance of less than a preset neighborhood radius are grouped into the same cluster. For each anomaly cluster, convex hull calculation or polygon fitting is performed to generate its initial boundary; then, morphological closing operations are used to fill the small internal holes, smooth the boundary, and form a continuous anomaly detection region polygon. During aggregation, adjacent regions with the same or similar anomaly types are merged first to form a panoramic region with semantic consistency, and a comprehensive anomaly type label and initial severity estimate are assigned to it.

[0010] As an optional implementation, in a first aspect of the present invention, determining the second acquisition task based on the spatial location information and area information of all abnormal sub-regions and window detection regions includes: Obtain the state vector of each drone in the drone swarm, the state vector including the current position and remaining energy; Calculate the estimated cost for each UAV to travel to each assigned task unit, forming a cost matrix; the estimated cost is calculated based on at least one of the following factors: flight distance, estimated flight time, energy consumption, and the matching weight of the task unit; Based on the cost matrix, a global optimization assignment algorithm is used to assign a corresponding execution drone to each task unit so as to minimize the total cost of all drones completing the assigned task. If the number of covering lines is equal to the order of the matrix, the optimal assignment scheme is determined from the zero element positions; if the number of covering lines is less than the order of the matrix, the matrix is ​​transformed to increase the number of zero elements, and the covering and judgment steps are repeated until the number of covering lines is equal to the order of the matrix.

[0011] As an optional implementation, in the first aspect of the present invention, the global optimization assignment algorithm based on the cost matrix includes: For each row of the cost matrix, subtract the minimum element value of that row; for each column of the resulting matrix, subtract the minimum element value of that column; cover all zero elements in the matrix with the minimum number of horizontal or vertical lines. The transformation of the matrix to increase the number of zero elements includes: finding the minimum value among the elements not covered by lines; subtracting the minimum value from all elements not covered by lines; and adding the minimum value to all elements covered by two lines.

[0012] A second aspect of this invention discloses a vegetation AI recognition and management system based on unmanned aerial vehicles (UAVs), comprising: First acquisition module: used to control the first UAV to perform the first acquisition task at the first flight altitude, acquire wide-area remote sensing data of the target detection area; and identify the wide-area remote sensing data to obtain all abnormal sub-regions and window detection areas within the target detection area; Task determination module: used to determine the second acquisition task based on the spatial location and area information of all abnormal sub-regions and window detection areas; The second acquisition module is used to control the second UAV to perform a second acquisition task at a second flight altitude lower than the first flight altitude, for all abnormal sub-regions and window detection areas to obtain corresponding fine remote sensing data. The first identification module is used to determine the vegetation status information of the abnormal sub-regions based on the fine remote sensing data, determine the canopy shading relationship in the window detection area based on the fine remote sensing data, and identify the potential understory area under the canopy cover. The second identification module is used to probabilistically infer the vegetation status of potential understory areas by fusing multi-source evidence, and generate a probability distribution map of understory vegetation that includes the inferred vegetation category and corresponding confidence level; wherein, the multi-source evidence includes direct observation data of window detection areas identified from the fine remote sensing data, partial exposure features of canopy edge areas, and correlation rules based on ecological prior knowledge; Result generation module: used to generate a comprehensive vegetation status report of the target detection area based on the wide-area status information of the wide-area remote sensing data, the refined identification results of the fine remote sensing data, and the inference information of the forest understory vegetation probability distribution map.

[0013] A third aspect of the present invention discloses an electronic device, comprising: a memory storing executable program code; a processor coupled to the memory; the processor calling the executable program code stored in the memory to execute the UAV-based vegetation AI recognition and management method disclosed in the first aspect of the present invention.

[0014] The fourth aspect of this invention discloses a computer-readable storage medium storing a computer program, wherein the computer program causes a computer to execute the UAV-based vegetation AI identification and management method disclosed in the first aspect of this invention.

[0015] Compared with the prior art, the embodiments of the present invention have the following beneficial effects: The drone-based vegetation AI recognition and management method in this embodiment first controls a first drone to acquire wide-area remote sensing data of the target detection area at a first flight altitude. This allows for a quick and comprehensive understanding of the overall area, identifying abnormal sub-regions and window detection areas, providing guidance for subsequent fine-grained detection. Then, a second drone is controlled at a lower second flight altitude to acquire fine-grained remote sensing data for these specific areas. This layered acquisition method ensures overall area coverage while enabling in-depth and detailed detection of key areas, improving the efficiency and accuracy of information acquisition. Attached Figure Description

[0016] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0017] Figure 1 This is a flowchart illustrating the vegetation AI recognition and management method based on drones disclosed in an embodiment of the present invention; Figure 2 This is a schematic diagram of the window region detection process disclosed in an embodiment of the present invention; Figure 3 This is a schematic diagram of a vegetation AI recognition and management system based on unmanned aerial vehicles (UAVs) provided in an embodiment of the present invention. Figure 4 This is a schematic diagram of the structure of an electronic device provided in an embodiment of the present invention. Detailed Implementation

[0018] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0019] It should be noted that the terms "first," "second," "third," "fourth," etc., in the specification and claims of this invention are used to distinguish different objects, not to describe a specific order. The terms "comprising" and "having," and any variations thereof, in the embodiments of this invention are intended to cover non-exclusive inclusion. Exemplarily, a process, method, system, product, or device that includes a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to these processes, methods, products, or devices.

[0020] Example 1 Please see Figure 1 , Figure 1 This is a flowchart illustrating the drone-based AI-powered vegetation identification and management method disclosed in this invention. The execution entity of the method described in this embodiment is composed of software and / or hardware. This entity can receive relevant information via wired or / or wireless means and can send certain instructions. It may also have processing and storage capabilities. This entity can control multiple devices, such as remote physical servers or cloud servers and related software, or local hosts or servers and related software that perform operations on devices located in a specific location. In some scenarios, multiple storage devices can also be controlled; these storage devices may be located in the same or different locations as the devices. Figure 1 As shown, this drone-based AI-based vegetation identification and management method includes the following steps: S101: Control the first UAV to perform the first data acquisition task at the first flight altitude, acquire wide-area remote sensing data of the target detection area; and identify the wide-area remote sensing data to obtain all abnormal sub-regions and window detection areas within the target detection area; S102: Determine the second acquisition task based on the spatial location and area information of all abnormal sub-regions and window detection areas; S103: Control the second UAV to perform a second acquisition task at a second flight altitude lower than the first flight altitude for all abnormal sub-regions and window detection areas to obtain corresponding fine remote sensing data; S104: Determine the vegetation status information of the abnormal sub-regions based on the fine remote sensing data, determine the canopy shading relationship in the window detection area based on the fine remote sensing data, and identify the potential understory area under the canopy cover; S105: By fusing multi-source evidence, the vegetation status of potential understory areas is probabilistically inferred, generating a probability distribution map of understory vegetation that includes the inferred vegetation category and corresponding confidence level; wherein, the multi-source evidence includes direct observation data of window detection areas identified from the fine remote sensing data, partial exposure features of canopy edge areas, and correlation rules based on ecological prior knowledge; S106: Based on the wide-area status information of the wide-area remote sensing data, the refined identification results of the fine remote sensing data, and the inference information of the forest understory vegetation probability distribution map, generate a comprehensive vegetation status report for the target detection area.

[0021] The solution of this invention, through data collection at different heights, not only obtains surface information of vegetation, but also combines it with subsequent analysis to obtain information such as vegetation status, canopy shading relationship and potential understory area information, so as to comprehensively understand the vegetation situation from multiple dimensions and provide a rich and accurate data foundation for subsequent vegetation management.

[0022] Specifically, this method uses multi-source evidence to probabilistically infer the vegetation status of potential understory areas. This multi-source evidence includes direct observation data from areas identified through high-resolution remote sensing, partial exposure features of canopy edges, and correlation rules based on prior ecological knowledge. This inference approach, integrating multiple information sources, fully considers the influence of different factors on understory vegetation status, making the inference results more scientific and reasonable. The generated understory vegetation probability distribution map accurately reflects different vegetation categories and their confidence levels, providing a more reliable basis for vegetation management decisions.

[0023] Specifically, during flight detection at the first altitude, if a large area remains unaffected (normal state), no action is taken to improve overall recognition speed. This recognition focuses on vegetation indices. However, if abnormal areas are identified, their locations are marked, and the abnormal areas within that region are aggregated to form a panoramic anomaly detection area. This panoramic anomaly detection area makes it easier to locate pests, diseases, fires, or other conditions. The corresponding status information is determined based on the accuracy of the drone's recognition at the first altitude. This status information includes levels one to four, where level one is normal and level four is abnormal. Level one drones do not require inspection, while level four drones must be inspected. Levels two and three require further assessment. The proportions of levels two and three can be adjusted based on the specific mission requirements of the level two drones.

[0024] More preferably, the step of determining the canopy shading relationship in the window detection area based on the refined remote sensing data, and identifying potential understory areas under the canopy cover, includes: S1041: Perform semantic segmentation on the fine remote sensing data to identify and segment canopy gaps; extract spectral and texture features within each window detection area, and use the first vegetation classification model to classify the vegetation within each window detection area to obtain the first set of observed species and their spatial locations. S1042: Identify the outline edges of each canopy object in the fine remote sensing data, and extract the edge transition zone of a preset width; analyze the image features within the edge transition zone, and identify the partially exposed understory vegetation features through the second vegetation classification model to obtain the second observation species set; S1043: Establish a spatial probability propagation model, using the species information in the first and second observed species sets as known evidence sources, and calculate the initial probability of each species existing at each location in the potential understory area based on the distance decay function and the canopy density decay coefficient. S1044: Call the ecological knowledge graph, which includes rules on species symbiosis, light requirements, and terrain preferences; constrain and adjust the initial probability based on the ecological knowledge graph to generate the probability distribution map of the understory vegetation.

[0025] The scheme of this invention extracts spectral and texture features within the detection area of ​​each window, and uses a first vegetation classification model to classify the vegetation, obtaining a first set of observed species and their spatial locations. Spectral and texture features can reflect the unique attributes of vegetation, and combined with the classification model, different species can be accurately identified.

[0026] In practice, the edge transition zone extraction is ingenious. It identifies the contour edges of each canopy object in the detailed remote sensing data and extracts an edge transition zone of a preset width. This operation focuses on the special region of the canopy edge because this region often provides some information about the understory vegetation, providing a targeted data range for further identification of understory vegetation characteristics.

[0027] The second classification model accurately identifies and analyzes image features within the edge transition zone. It then identifies partially exposed understory vegetation features using a second vegetation classification model, resulting in a second observation species set. This second classification model is specifically trained for the features of the edge transition zone, enabling it to more effectively identify understory vegetation exposed at the canopy edge. This complements the deficiencies of the first observation species set, making the identification of understory vegetation more comprehensive.

[0028] A spatial probability propagation model was established, using species information from the first and second observation sets as known evidence sources. Based on the distance decay function and the canopy density decay coefficient, the initial probability of each species existing at each location in the potential understory area was calculated. The distance decay function considers the distance relationship between species distribution and known observation points, while the canopy density decay coefficient considers the shading effect of the canopy on understory vegetation. This comprehensive calculation method, which considers multiple factors, makes the initial probability calculation more scientific and reasonable, and can reflect the probability of different species existing at different locations.

[0029] In this embodiment of the invention, the initial probabilities are constrained and adjusted based on an ecological knowledge graph, which makes the generated understory vegetation probability distribution map more consistent with ecological reality. For example, based on species association relationships, if a species usually coexists with another species, this relationship will be considered when adjusting the probability; based on light requirements and terrain preference rules, the distribution probability of different species in different locations will be further optimized, so that the probability distribution map can more accurately reflect the true state of vegetation in potential understory areas.

[0030] More preferably, the wide-area remote sensing data is identified to obtain all window detection regions within the target detection area, including: The wide-area remote sensing data is preprocessed to generate a canopy cover binarized mask image; The seed connected region growing algorithm is used to mark and segment the non-canopy region in the canopy cover binarized mask image to identify independent window patches; Calculate the area, perimeter, minimum bounding rectangle, and shape complexity index of each window patch; Alternatively, the wide-area remote sensing data can be identified to obtain all window detection regions within the target detection area, including: Identify all window detection regions within the target area and calculate the window area and geometric features of each window detection region; Perform a second acquisition task on all window detection areas to obtain the corresponding detailed remote sensing data, including: Based on the window area of ​​the target detection region and the geometric features, the corresponding second UAV is dynamically determined and assigned to perform the second data acquisition task at the second flight altitude. For window detection areas with a window area larger than a preset window threshold, a single UAV performs a detailed grid survey. For multiple window detection areas with a window area smaller than the preset window threshold and discretely distributed, the corresponding acquisition points are determined based on the positions between each window detection area so that the second UAV can acquire multiple window detection areas at the second flight altitude; and the corresponding second acquisition task is performed to acquire the corresponding fine remote sensing data.

[0031] Based on the window area and geometric features of the target detection region, a corresponding second UAV is dynamically determined and assigned to perform the second data acquisition task at a second flight altitude. This dynamic assignment method fully considers the differences between different window detection regions, and can rationally allocate UAV resources according to the actual situation of the window, avoiding resource waste and improving the targeting and efficiency of the data acquisition task.

[0032] For window detection areas with a window area larger than a preset window threshold, a single drone performs a detailed grid survey. This grid survey ensures comprehensive and meticulous coverage of the area, leaving no important information unaccounted for. Since large window areas typically contain more vegetation information, requiring more refined data collection, the grid survey method meets this need, providing high-quality data for subsequent accurate analysis of vegetation status.

[0033] For multiple window detection areas with a window area smaller than a preset window threshold and a discrete distribution, corresponding acquisition points are determined based on the positions between each window detection area, enabling the second UAV to acquire multiple window detection areas at the second flight altitude. This method optimizes the UAV's flight path, reduces flight distance and time, and improves acquisition efficiency. Simultaneously, by rationally planning acquisition points, effective coverage of all small window areas is ensured, preventing the importance of small windows from being overlooked.

[0034] More preferably, the step of determining the canopy shading relationship in the window detection area based on the refined remote sensing data, and identifying potential understory areas under the canopy cover, includes: The detailed remote sensing data is analyzed to determine the causal type of windows in the window detection area, including a first type of window formed by gaps between adjacent trees and a second type of window formed by openings in the canopy of a single tree. Based on the aforementioned cause types, different spatial reasoning models are adopted. For the first type of window, it is treated as an independent observation window, and its internal vegetation is directly classified. For the second type of window, it is associated with the canopy model of the tree to which it belongs, and its edge features are analyzed to infer the overall canopy health status and lower light transmission conditions of the tree. By integrating direct classification and inference results, a forest understory vegetation status map containing vegetation categories and corresponding confidence levels is generated.

[0035] The scheme of this invention analyzes refined remote sensing data to determine the cause type of windows in the window detection area, clearly distinguishing between the first type of windows formed by gaps between adjacent trees and the second type of windows formed by openings in the canopy of a single tree. This detailed classification of causes allows for a deeper understanding of the essential reasons for window formation, because windows of different causes differ in spatial characteristics, surrounding vegetation distribution, and impact on understory vegetation, providing an accurate basis for subsequently employing different analysis methods.

[0036] By utilizing high-resolution remote sensing data for analysis, the information contained within the data can be fully extracted. High-resolution remote sensing data has high resolution and rich detail. Through in-depth analysis of this data, the cause of the window can be accurately determined, avoiding subsequent analysis errors caused by inaccurate cause determination, and improving the accuracy and reliability of the entire analysis process.

[0037] Specifically, for the first type of window, it is treated as an independent observation window, and the vegetation within it is directly classified. Since the first type of window is formed by gaps between adjacent trees, the vegetation within it is relatively independent and unaffected by the complex canopy structure of the trees. Direct classification can quickly and accurately identify the vegetation category within the window, improving analysis efficiency. At the same time, the direct classification method is simple and straightforward, reducing errors caused by complex models and ensuring the accuracy of the classification results.

[0038] For the second type of window, it is associated with the canopy model of the tree to analyze its edge features and infer the overall canopy health and understory light transmission conditions of the tree. The second type of window is closely related to the canopy of the tree to which it belongs. By associating it with the canopy model, the influence of factors such as the tree's overall structure and growth status on the window can be comprehensively considered. Analyzing the edge features can obtain more detailed information about the tree canopy, thus more accurately inferring the overall canopy health and understory light transmission conditions, which is of great significance for understanding the growth environment of understory vegetation.

[0039] More preferably, before controlling the first UAV to perform the first data collection task at the first flight altitude, the method further includes: Obtain the boundary range of the target area and multiple pre-divided or automatically generated management plot units within it; Based on preliminary analysis of historical data, land cover maps or initial remote sensing data, each management plot unit is assigned a corresponding attribute label, which includes one or more of the following: vegetation type, functional zoning, management priority, estimated window density, and historical pest and disease incidence rate. Based on the attribute tags, an initial set of drone operation parameter suggestions is generated for each of the managed plot units. The set of operation parameter suggestions includes recommended flight altitude, sensor type, flight path overlap rate, and whether to enable cluster collaboration mode.

[0040] In practice, the boundary of the target area and multiple pre-divided or automatically generated management plots within it are obtained, clarifying the specific scope and internal structure of the target area. This division method makes the management of the entire target area more orderly and refined, providing a basic framework for developing personalized operation plans for different plots and avoiding blind and generalized operations.

[0041] Whether pre-divided or automatically generated by algorithms, management plot units can be flexibly adjusted according to the actual situation of the target area. For example, for areas with complex terrain and uneven vegetation distribution, the algorithm can automatically generate suitable plot units based on factors such as terrain and vegetation, ensuring that each plot has relatively consistent characteristics, which facilitates subsequent targeted operation planning.

[0042] Based on preliminary analysis of historical data, land cover maps, or initial remote sensing data, each management plot unit is assigned corresponding attribute labels. These attribute labels cover a variety of information, including vegetation type, functional zoning, management priority, estimated window density, and historical pest and disease incidence rates. These attribute labels provide a comprehensive understanding of the characteristics and conditions of each management plot unit, offering rich evidence for developing scientific and reasonable operational plans.

[0043] Different attribute tags reflect the different needs and characteristics of each plot of land. For example, functional zoning affects the focus of operations and safety requirements; management priorities determine the order of operations; estimated window density helps plan drone flight routes and collection points; and historical pest and disease incidence rates provide important references for pest and disease monitoring and control. Based on these attribute tags, personalized operation plans can be developed for each plot of land, improving the targeting and effectiveness of operations.

[0044] More preferably, the wide-area remote sensing data is identified to obtain all anomalous sub-regions within the target detection area, including: The wide-area remote sensing data is identified to obtain abnormal pixels or abnormal patches within the target detection area. Based on their geographic coordinates, a density-based spatial clustering algorithm is used to group them, and abnormal points with a spatial distance of less than a preset neighborhood radius are grouped into the same cluster. For each anomaly cluster, convex hull calculation or polygon fitting is performed to generate its initial boundary; then, morphological closing operations are used to fill the small internal holes, smooth the boundary, and form a continuous anomaly detection region polygon. During aggregation, adjacent regions with the same or similar anomaly types are merged first to form a panoramic region with semantic consistency, and a comprehensive anomaly type label and initial severity estimate are assigned to it.

[0045] This invention identifies anomalous pixels or patches within a target detection area by performing wide-area remote sensing data identification. This step enables the precise location of potentially abnormal areas from massive amounts of remote sensing data. Through specific identification algorithms, pixels or patches that differ from normal conditions can be effectively filtered out.

[0046] This invention employs a density-based spatial clustering algorithm based on the geographic coordinates of anomalies or patches to group them into clusters. Anomalies with a spatial distance less than a preset neighborhood radius are grouped together. This clustering method can reasonably group anomalies according to their spatial distribution density, without being limited by the shape or size of the anomalies. It can effectively aggregate adjacent or nearby anomalies to form meaningful anomaly clusters, avoiding isolated processing of anomalies and improving the accuracy and completeness of anomaly area identification.

[0047] For each anomaly cluster, convex hull calculation or polygon fitting is performed to generate its initial boundary. Convex hull calculation finds the smallest convex polygon containing all points in the anomaly cluster, while polygon fitting fits a polygon that more closely approximates the actual boundary based on the shape characteristics of the anomaly cluster. Both methods can determine a rough boundary range for the anomaly cluster, providing a foundation for subsequent boundary optimization.

[0048] Morphological closing operations fill in small internal holes and smooth the boundaries, forming continuous polygonal anomaly detection regions. Morphological closing is a commonly used image processing technique that first dilates the image to fill in small internal holes, then performs erosion to smooth the boundaries. After closing processing, the boundaries of the anomaly detection region are more continuous and smooth, more accurately reflecting the actual extent of the anomaly and improving the accuracy of anomaly identification.

[0049] During aggregation, adjacent areas with the same or similar anomaly types are preferentially merged to form a panoramic region with semantic consistency. This merging method not only considers the spatial relationship of anomaly areas but also incorporates information about the anomaly type, making the merged panoramic region more semantically consistent and more in line with actual surface features and anomalies. For example, merging adjacent vegetation anomaly areas caused by pests and diseases can more comprehensively reflect the spread and impact of pests and diseases, providing more valuable information for subsequent analysis and decision-making.

[0050] More preferably, the step of determining the second acquisition task based on the spatial location and area information of all abnormal sub-regions and window detection regions includes... Obtain the state vector of each drone in the drone swarm, the state vector including the current position and remaining energy; Calculate the estimated cost for each UAV to travel to each assigned task unit, forming a cost matrix; the estimated cost is calculated based on at least one of the following factors: flight distance, estimated flight time, energy consumption, and the matching weight of the task unit; Based on the cost matrix, a global optimization assignment algorithm is used to assign a corresponding execution drone to each task unit so as to minimize the total cost of all drones completing the assigned task. If the number of covering lines is equal to the order of the matrix, the optimal assignment scheme is determined from the zero element positions; if the number of covering lines is less than the order of the matrix, the matrix is ​​transformed to increase the number of zero elements, and the covering and judgment steps are repeated until the number of covering lines is equal to the order of the matrix.

[0051] The estimated cost for each drone to travel to each assigned task unit is calculated. This estimated cost is based on at least one of the following factors: flight distance, estimated flight time, energy consumption, and the task unit's matching weight. Flight distance and estimated flight time directly affect task execution efficiency, while energy consumption relates to the drone's endurance and the reliability of task completion. The task unit's matching weight considers the compatibility between the drone and the task; for example, some drones may be better suited for specific types of tasks. By comprehensively considering these factors, the cost of each drone executing each task unit can be more comprehensively and accurately assessed, resulting in a more reasonable cost matrix.

[0052] Based on the cost matrix, a global optimization assignment algorithm is used to allocate appropriate execution drones to each task unit, minimizing the total cost for all drones to complete their assigned tasks. This algorithm considers the relationships between all drones and task units from a holistic perspective, seeking the optimal task allocation scheme rather than merely pursuing local optima. This fully utilizes the resources of the drone swarm, improving task execution efficiency and quality while reducing overall task execution costs. When allocating tasks, the algorithm comprehensively considers the state of each drone and the characteristics of the task unit, avoiding situations where some drones are overloaded while others are underloaded. By minimizing the total cost, the algorithm can rationally allocate tasks, ensuring that each drone can undertake an appropriate amount of work within its capabilities, thus improving the fairness and rationality of task allocation.

[0053] More preferably, a global optimization assignment algorithm based on the cost matrix includes: For each row of the cost matrix, subtract the minimum element value of that row; for each column of the resulting matrix, subtract the minimum element value of that column; cover all zero elements in the matrix with the minimum number of horizontal or vertical lines. The transformation of the matrix to increase the number of zero elements includes: finding the minimum value among the elements not covered by lines; subtracting the minimum value from all elements not covered by lines; and adding the minimum value to all elements covered by two lines.

[0054] The specific operational steps for transforming the cost matrix to increase the number of zero elements are detailed. First, the minimum value is found among the elements not covered by lines. Then, this minimum value is subtracted from all elements not covered by lines, and finally, it is added to all elements covered by two lines. These steps effectively adjust the element values ​​in the cost matrix, increasing the number of zero elements and thus providing more possibilities for finding the optimal assignment scheme.

[0055] To improve the convergence speed of the algorithm, increasing the number of zero elements simplifies the process of covering all zero elements in the matrix with horizontal or vertical lines. This reduces the number of line count checks and matrix transformations, thus increasing the convergence speed. This allows the algorithm to find the optimal assignment scheme more quickly, reducing task allocation time and improving the response speed and task execution efficiency of the drone swarm.

[0056] The solutions in this embodiment of the invention also include: When the number of task units exceeds the number of available drones, the method also includes: Group task units into task unit groups equal to the number of available drones, based on spatial location or task priority. Treat each task unit group as a whole task and perform global optimized assignment; Once a drone is assigned to a task unit group, its execution sequence is planned within that group according to the shortest path principle.

[0057] In the actual implementation, targeted early warnings can also be issued. After identifying abnormal areas, further steps such as dynamic ecological risk assessment and early warning zone delineation can be carried out. Based on the attributes of the abnormal area, determine its potential diffusion type; the diffusion type includes disease and biological transmission, pest migration, fire spread or physiological stress diffusion; match or call a corresponding spatiotemporal diffusion prediction model for the diffusion type; Using the currently identified abnormal areas as the diffusion source, input the current and predicted environmental driving factors, including wind direction and speed, temperature, humidity, terrain slope, and vegetation connectivity. Run the space-time diffusion prediction model to simulate the diffusion range and probability of the abnormal state within a preset time period in the future. Based on the simulation results, one or more risk warning areas are delineated; the risk warning areas include high-probability spread areas and key blocking point areas; among them, the high-probability spread areas are the areas that the abnormal state may directly spread to in the model prediction, and the key blocking point areas are the locations that have strategic value in terms of terrain or vegetation structure for preventing spread.

[0058] The AI ​​model used in the method of this invention adopts a two-tiered targeted modeling architecture that is both task-driven and data-driven: The first-level lightweight screening model is designed specifically for processing multispectral images acquired by the first UAV. Its input is multiple pre-computed vegetation index maps or raw multispectral bands, and its output is the anomaly probability for each pixel or image patch. Its network structure is a lightweight fully convolutional network that has been deeply compressed and pruned to meet the real-time processing requirements of airborne devices. Its training objective is to maximize the ability to distinguish between healthy and abnormal states, rather than precise species or etiology classification. The second-level refined diagnostic model is designed specifically for processing high-resolution visible light images acquired by the second UAV. Its input is a high-resolution RGB image patch, and it can selectively fuse the anomaly probability map of the corresponding region provided by the first-level model as attention guidance. Its output includes at least the specific vegetation species, pest and disease types, and severity levels. Its network structure is a deep learning network containing a feature pyramid to simultaneously capture detailed textures and high-level semantic features. Its training objective is to achieve accurate localization of disease patches or instance segmentation of individual plants based on accurate classification.

[0059] The second-level refined diagnostic model adopts a context-aware multi-task learning network architecture: the network has a shared feature extraction backbone network; on the backbone network, three task-specific sub-network heads are connected in parallel: a classification head for classifying vegetation and pests, a detection / segmentation head for locating diseased areas or individual plants, and a regression head for assessing disease severity or vegetation vitality. During training, the loss functions of the classification head, detection / segmentation head, and regression head are jointly optimized, enabling the model to learn associated morphological, textural, and spatial context features from high-resolution visible light images and simultaneously output multiple pieces of information required for diagnosis.

[0060] The method in this embodiment of the invention further includes knowledge distillation and co-evolution steps in a two-level model: After the second-level refined diagnostic model has been trained and matured, it will be used as a teacher model. Design a specific knowledge distillation training process, so that the first-level lightweight screening model acts as a student model, which not only learns to distinguish between normal and abnormal basic tasks, but also tries to learn to imitate the soft classification probability distribution of abnormal types of the teacher model (even if the student model does not output these types in the end). This distillation process compresses and transfers the fine-grained discriminative knowledge inherent in the second-level model to the first-level model, thereby improving the accuracy of anomaly screening and the initial identification of anomaly types without increasing the complexity of the first-level model. During deployment, the second-level refined diagnostic model dynamically loads different specialized sub-models based on the first-level screening results for the target area.

[0061] Establish an expert model library containing specialized diagnostic models trained and optimized for different major vegetation types (such as fruit trees, broad-leaved forests, coniferous forests, and lawns) or different major diseases (such as fungal, bacterial, and insect pests). When the first-level screening indicates that an abnormal area is dominated by a certain vegetation type or is suspected of having a certain disease, the corresponding specialized expert diagnostic model is dynamically loaded onto the edge computing device on the second UAV responsible for detailed investigation of the area to replace the general diagnostic model, thereby achieving the highest diagnostic accuracy for this specific task.

[0062] The drone-based vegetation AI recognition and management method in this embodiment first controls a first drone to acquire wide-area remote sensing data of the target detection area at a first flight altitude. This allows for a quick and comprehensive understanding of the overall area, identifying abnormal sub-regions and window detection areas, providing guidance for subsequent fine-grained detection. Then, a second drone is controlled at a lower second flight altitude to acquire fine-grained remote sensing data for these specific areas. This layered acquisition method ensures overall area coverage while enabling in-depth and detailed detection of key areas, improving the efficiency and accuracy of information acquisition.

[0063] Example 2 Please see Figure 4 , Figure 4 This is a schematic diagram of the structure of the drone-based vegetation AI recognition and management system disclosed in an embodiment of the present invention. Figure 4 As shown, the drone-based vegetation AI recognition and management system may include: First acquisition module 21: used to control the first UAV to perform the first acquisition task at the first flight altitude, acquire wide-area remote sensing data of the target detection area; and identify the wide-area remote sensing data to obtain all abnormal sub-regions and window detection areas within the target detection area; Task determination module 22: used to determine the second acquisition task based on the spatial location and area information of all abnormal sub-regions and window detection areas; Second acquisition module 23: Used to control the second UAV to perform a second acquisition task at a second flight altitude lower than the first flight altitude for all abnormal sub-regions and window detection areas to obtain corresponding fine remote sensing data; First identification module 24: used to determine vegetation status information of abnormal sub-regions based on the fine remote sensing data and to determine the canopy shading relationship in the window detection area based on the fine remote sensing data, and to identify potential understory areas under the canopy cover; The second identification module 25 is used to make a probabilistic inference on the vegetation status of potential understory areas by fusing multi-source evidence, and generate a probability distribution map of understory vegetation that includes the inferred vegetation category and the corresponding confidence level; wherein, the multi-source evidence includes direct observation data of the window detection area identified from the fine remote sensing data, partial exposure features of the canopy edge area, and correlation rules based on ecological prior knowledge. Result generation module 26: used to generate a comprehensive vegetation status report of the target detection area based on the wide-area status information of the wide-area remote sensing data, the refined identification results of the fine remote sensing data, and the inference information of the forest understory vegetation probability distribution map.

[0064] The drone-based vegetation AI recognition and management method in this embodiment first controls a first drone to acquire wide-area remote sensing data of the target detection area at a first flight altitude. This allows for a quick and comprehensive understanding of the overall area, identifying abnormal sub-regions and window detection areas, providing guidance for subsequent fine-grained detection. Then, a second drone is controlled at a lower second flight altitude to acquire fine-grained remote sensing data for these specific areas. This layered acquisition method ensures overall area coverage while enabling in-depth and detailed detection of key areas, improving the efficiency and accuracy of information acquisition.

[0065] Example 3 Please see Figure 4 , Figure 4 This is a schematic diagram of the structure of an electronic device disclosed in an embodiment of the present invention. The electronic device can be a computer, a server, etc. Of course, in certain cases, it can also be a mobile phone, tablet computer, monitoring terminal, or other smart device, as well as an image acquisition device with processing capabilities. Figure 4 As shown, the electronic device may include: Memory 510 storing executable program code; Processor 520 coupled to memory 510; The processor 520 calls the executable program code stored in the memory 510 to execute some or all of the steps in the drone-based vegetation AI recognition and management method in Embodiment 1.

[0066] This invention discloses a computer-readable storage medium storing a computer program that causes a computer to perform some or all of the steps in the drone-based vegetation AI identification and management method of Embodiment 1.

[0067] This invention also discloses a computer program product, wherein when the computer program product is run on a computer, the computer performs some or all of the steps in the drone-based vegetation AI identification and management method in Embodiment 1.

[0068] This invention also discloses an application publishing platform, which is used to publish computer program products. When the computer program products are run on a computer, the computer executes some or all of the steps in the drone-based vegetation AI identification and management method in Embodiment 1.

[0069] In various embodiments of the present invention, it should be understood that the sequence number of each process does not necessarily imply the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of the present invention.

[0070] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; they can be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0071] Furthermore, the functional units in the various embodiments of the present invention can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit.

[0072] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-accessible memory. Based on this understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a memory and includes several requests to cause a computer device (which can be a personal computer, server, or network device, specifically a processor in the computer device) to execute some or all of the steps of the methods described in the various embodiments of the present invention.

[0073] In the embodiments provided by this invention, it should be understood that "B corresponding to A" means that B is associated with A, and B can be determined based on A. However, it should also be understood that determining B based on A does not mean determining B solely based on A; B can also be determined based on A and / or other information.

[0074] Those skilled in the art will understand that some or all of the steps in the various methods of the embodiments described can be implemented by a program instructing related hardware. This program can be stored in a computer-readable storage medium, including read-only memory (ROM), random access memory (RAM), programmable read-only memory (PROM), erasable programmable read-only memory (EPROM), one-time programmable read-only memory (OTPROM), electrically-Erasable Programmable Read-Only Memory (EEPROM), compact disc read-only memory (CD-ROM) or other optical disc storage, disk storage, magnetic tape storage, or any other computer-readable medium capable of carrying or storing data.

[0075] The above provides a detailed description of the drone-based AI-based vegetation recognition and management method, system, electronic device, and storage medium disclosed in the embodiments of the present invention. Specific examples have been used to illustrate the principles and implementation methods of the present invention. The descriptions of the above embodiments are only for the purpose of helping to understand the method and core ideas of the present invention. At the same time, for those skilled in the art, there will be changes in the specific implementation methods and application scope based on the ideas of the present invention. Therefore, the content of this specification should not be construed as a limitation of the present invention.

Claims

1. A method for vegetation AI recognition and management based on drones, characterized in that, include: Control the first UAV to perform the first data acquisition task at the first flight altitude, and acquire wide-area remote sensing data of the target detection area; The wide-area remote sensing data is then identified to obtain all abnormal sub-regions and window detection regions within the target detection area; The second acquisition task is determined based on the spatial location and area information of all abnormal sub-regions and window detection regions; the windows of the window detection region include a first type of window formed by the gaps between adjacent trees and a second type of window formed by the openings in the canopy of a single tree itself; Control the second UAV to perform a second acquisition task at a second flight altitude lower than the first flight altitude, and perform a second acquisition task on all abnormal sub-regions and window detection areas to obtain corresponding fine remote sensing data; Based on the fine remote sensing data, vegetation status information of abnormal sub-regions is determined, and canopy shading relationships in the window detection area are determined based on the fine remote sensing data, and potential understory areas under canopy cover are identified. By fusing multi-source evidence, a probabilistic inference of the vegetation status of potential understory areas is made, generating a probability distribution map of understory vegetation containing inferred vegetation categories and corresponding confidence levels. The multi-source evidence includes direct observation data of window detection areas identified from the refined remote sensing data, partial exposure features of canopy edge areas, and association rules based on prior ecological knowledge. The process of probabilistically inferring the vegetation status of potential understory areas by fusing multi-source evidence to generate a probability distribution map of understory vegetation containing inferred vegetation categories and corresponding confidence levels includes: Semantic segmentation is performed on the refined remote sensing data to identify and segment canopy gaps; spectral and texture features are extracted from each window detection area, and the vegetation in each window detection area is classified using a first vegetation classification model to obtain a first set of observed species and their spatial locations. Identify the outline edges of each canopy object in the refined remote sensing data, and extract the edge transition zone of a preset width; analyze the image features within the edge transition zone, and identify the partially exposed understory vegetation features through a second vegetation classification model to obtain a second set of observed species; A spatial probability propagation model is established, using the species information in the first and second observed species sets as known evidence sources. Based on the distance decay function and the canopy density decay coefficient, the initial probability of each species existing at each location in the potential understory area is calculated. An ecological knowledge graph is invoked, which includes rules on species symbiosis, light requirements, and terrain preferences. Based on the ecological knowledge graph, the initial probability is constrained and adjusted to generate the probability distribution map of the forest understory vegetation. Based on the wide-area status information of the wide-area remote sensing data, the vegetation status information of the abnormal sub-regions of the fine remote sensing data, and the inferred information of the forest understory vegetation probability distribution map, a comprehensive vegetation status report of the target detection area is generated.

2. The vegetation AI recognition and management method based on unmanned aerial vehicles as described in claim 1, characterized in that, The wide-area remote sensing data is identified to obtain all window detection regions within the target detection area, including: The wide-area remote sensing data is preprocessed to generate a canopy cover binarized mask image; The seed connected region growing algorithm is used to mark and segment the non-canopy region in the canopy cover binarized mask image to identify independent window patches; Calculate the area, perimeter, minimum bounding rectangle, and shape complexity index of each window patch; Alternatively, the wide-area remote sensing data can be identified to obtain all window detection regions within the target detection area, including: Identify all window detection regions within the target detection region, and calculate the window area and geometric features of each window detection region; Perform a second acquisition task on all window detection areas to obtain the corresponding detailed remote sensing data, including: Based on the window area of ​​the target detection region and the geometric features, the corresponding second UAV is dynamically determined and assigned to perform the second data acquisition task at the second flight altitude. For window detection areas with a window area larger than a preset window threshold, a single UAV performs a detailed grid survey. For multiple window detection areas with a window area smaller than the preset window threshold and discretely distributed, the corresponding acquisition points are determined based on the positions between each window detection area so that the second UAV can acquire multiple window detection areas at the second flight altitude; and the corresponding second acquisition task is performed to acquire the corresponding fine remote sensing data.

3. The vegetation AI recognition and management method based on unmanned aerial vehicles as described in claim 1, characterized in that, The step of determining the canopy shading relationship in the window detection area based on the refined remote sensing data, and identifying potential understory areas under the canopy cover, includes: Analyze the detailed remote sensing data to determine the causal type of windows in the window detection area; Based on the aforementioned cause types, different spatial reasoning models are adopted. For the first type of window, it is treated as an independent observation window, and its internal vegetation is directly classified. For the second type of window, it is associated with the canopy model of the tree to which it belongs, and its edge features are analyzed to infer the overall canopy health status and lower light transmission conditions of the tree. By integrating the direct classification with the inference results of the tree's overall canopy health and understory light transmission conditions, a forest understory vegetation status map containing vegetation category and corresponding confidence level is generated.

4. The vegetation AI recognition and management method based on unmanned aerial vehicles as described in claim 1, characterized in that, Before controlling the first UAV to perform the first data acquisition task at the first flight altitude, the method further includes: Obtain the boundary range of the target detection area and multiple pre-divided or automatically generated management plot units within it; Based on preliminary analysis of historical data, land cover maps or initial remote sensing data, each management plot unit is assigned a corresponding attribute label, which includes one or more of the following: vegetation type, functional zoning, management priority, estimated window density, and historical pest and disease incidence rate. Based on the attribute tags, an initial set of drone operation parameter suggestions is generated for each of the managed plot units. The set of operation parameter suggestions includes recommended flight altitude, sensor type, flight path overlap rate, and whether to enable cluster collaboration mode.

5. The vegetation AI recognition and management method based on unmanned aerial vehicles as described in claim 1, characterized in that, The wide-area remote sensing data is identified to obtain all anomalous sub-regions within the target detection area, including: The wide-area remote sensing data is identified to obtain abnormal pixels within the target detection area. Based on their geographic coordinates, a density-based spatial clustering algorithm is used to group them, and abnormal pixels with a spatial distance less than a preset neighborhood radius are grouped into the same cluster. For each anomaly cluster, convex hull calculation or polygon fitting is performed to generate its initial boundary; then, morphological closing operations are used to fill the small internal holes, smooth the boundary, and form a continuous anomaly detection region polygon. During aggregation, adjacent regions with the same or similar anomaly types are merged first to form a panoramic region with semantic consistency, and a comprehensive anomaly type label and initial severity estimate are assigned to it.

6. The vegetation AI recognition and management method based on unmanned aerial vehicles as described in claim 1, characterized in that, The second acquisition task is determined based on the spatial location and area information of all abnormal sub-regions and window detection areas, including: Obtain the state vector of each drone in the drone swarm, the state vector including the current position and remaining energy; Calculate the estimated cost for each UAV to travel to each assigned task unit, forming a cost matrix; the estimated cost is calculated based on at least one of the following factors: flight distance, estimated flight time, energy consumption, and the matching weight of the task unit; Based on the cost matrix, a global optimization assignment algorithm is used to assign a corresponding execution drone to each task unit so as to minimize the total cost of all drones completing the assigned task. The global optimization assignment algorithm includes: subtracting the minimum element value of each row of the cost matrix; subtracting the minimum element value of each column of the resulting matrix; and covering all zero elements in the matrix with the minimum number of horizontal or vertical lines. If the number of covering lines is equal to the order of the matrix, the optimal assignment scheme is determined from the zero element positions; if the number of covering lines is less than the order of the matrix, the matrix is ​​transformed to increase the number of zero elements, and the covering and judgment steps are repeated until the number of covering lines is equal to the order of the matrix.

7. The vegetation AI recognition and management method based on unmanned aerial vehicles as described in claim 6, characterized in that, The transformation of the matrix to increase the number of zero elements includes: finding the minimum value among the elements not covered by lines; subtracting the minimum value from all elements not covered by lines; and adding the minimum value to all elements covered by two lines.

8. A vegetation AI recognition and management system based on drones, characterized in that, include: First acquisition module: Used to control the first UAV to perform the first acquisition task at the first flight altitude and acquire wide-area remote sensing data of the target detection area; The wide-area remote sensing data is then identified to obtain all abnormal sub-regions and window detection regions within the target detection area; Task determination module: used to determine the second acquisition task based on the spatial location and area information of all abnormal sub-regions and window detection areas; the windows of the window detection area include a first type of window formed by the gaps between adjacent trees and a second type of window formed by the openings in the canopy of a single tree; The second acquisition module is used to control the second UAV to perform a second acquisition task at a second flight altitude lower than the first flight altitude, for all abnormal sub-regions and window detection areas to obtain corresponding fine remote sensing data. The first identification module is used to determine the vegetation status information of the abnormal sub-regions based on the fine remote sensing data, determine the canopy shading relationship in the window detection area based on the fine remote sensing data, and identify the potential understory area under the canopy cover. The second identification module is used to probabilistically infer the vegetation status of potential understory areas by fusing multi-source evidence, generating a probability distribution map of understory vegetation containing inferred vegetation categories and corresponding confidence levels. The multi-source evidence includes direct observation data of window detection areas identified from the refined remote sensing data, partial exposure features of canopy edge areas, and association rules based on prior ecological knowledge. The process of probabilistically inferring the vegetation status of potential understory areas by fusing multi-source evidence to generate a probability distribution map of understory vegetation containing inferred vegetation categories and corresponding confidence levels includes: Semantic segmentation is performed on the refined remote sensing data to identify and segment canopy gaps; spectral and texture features are extracted from each window detection area, and the vegetation in each window detection area is classified using a first vegetation classification model to obtain a first set of observed species and their spatial locations. Identify the outline edges of each canopy object in the refined remote sensing data, and extract the edge transition zone of a preset width; analyze the image features within the edge transition zone, and identify the partially exposed understory vegetation features through a second vegetation classification model to obtain a second set of observed species; A spatial probability propagation model is established, using the species information in the first and second observed species sets as known evidence sources. Based on the distance decay function and the canopy density decay coefficient, the initial probability of each species existing at each location in the potential understory area is calculated. An ecological knowledge graph is invoked, which includes rules on species symbiosis, light requirements, and terrain preferences. Based on the ecological knowledge graph, the initial probability is constrained and adjusted to generate the probability distribution map of the forest understory vegetation. Result generation module: used to generate a comprehensive vegetation status report of the target detection area based on the wide-area status information of the wide-area remote sensing data, the vegetation status information of the abnormal sub-regions of the fine remote sensing data, and the inference information of the forest understory vegetation probability distribution map.

9. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program, wherein the computer program causes a computer to perform the drone-based vegetation AI identification and management method according to any one of claims 1 to 7.

Images