Remote sensing image analysis-based corn northern leaf blight intelligent identification method and system

Abstract

The invention discloses a corn northern leaf blight intelligent identification method and system based on remote sensing image analysis, and belongs to the field of disease intelligent identification, and the method comprises the following steps: generating a flight route set according to the boundary information of a to-be-monitored field, obtaining a remote sensing image of a to-be-monitored area, and analyzing the remote sensing image to obtain a determined suspected point location; a region supplementary flight plan is generated by taking the determined suspected point location as a core, meanwhile, the severity index and the lower leaf confidence coefficient of a leaf region are calculated, and a disease distribution result of the whole field range is generated; and performing risk level label output on the field according to a disease distribution result, performing influence evaluation on a disease coverage area, and outputting a next round of inspection suggestion to an application end for display and management according to an influence evaluation result, thereby achieving the purposes of reducing early-stage missing detection conditions, improving result performability and improving inspection efficiency. The problem that in the prior art, in the early stage, the disease recognition accuracy is not high due to the fact that the remote sensing image cannot observe the lower leaves sufficiently is solved.

Description

Technical Field

[0001] This invention relates to the field of intelligent disease identification technology, and in particular to an intelligent identification method and system for maize leaf spot based on remote sensing image analysis. Background Technology

[0002] Maize leaf spot primarily damages leaves, significantly reducing leaf area and photosynthetic capacity in severe cases, thus affecting yield and quality. For large-scale field control of maize leaf spot, remote sensing images can generate consistent monitoring results over a wider area. By acquiring crop canopy images through remote sensing carriers such as drones or satellites and combining them with spectral information to construct characteristics that characterize plant health, deep learning models can be used to identify the disease and classify its severity. The results are then rasterized and output as a disease distribution map to support zonal application of pesticides, patrol guidance, and evaluation of control effects. This improves the efficiency of large-scale, rapid, and standardized field patrols and provides location and zoning basis for plant protection operations.

[0003] For example, Chinese Invention Patent CN116087118B discloses a device and system for identifying corn leaf blight using hyperspectral remote sensing. The device includes a drone, a hyperspectral remote sensing instrument connected to the drone, and a corn leaf clustering module. The drone carries the hyperspectral remote sensing instrument to a set altitude above the corn plantation to collect leaf data of the corn to be tested. The corn leaf clustering module is used to cluster the leaves of the corn to be tested sequentially from top to bottom. In this invention, when the drone carrying the hyperspectral remote sensing instrument collects data from the corn plant at high altitude, the corn leaf clustering mechanism sequentially clusters the leaves of the corn plant from top to bottom, enabling the hyperspectral remote sensing instrument to efficiently collect data from the lower leaves of the corn plant, ensuring that all leaf data of the corn plant can be collected and reducing blind spots in data collection.

[0004] In existing technologies, when acquiring canopy surface information using a top-view approach, limitations imposed by observation geometry and canopy occlusion exist. Top-view images naturally tend to reflect the condition of upper leaves. However, in the field spread of maize leaf spot disease, early lesions often appear on lower leaves, resulting in very weak or even absent visible phenotypes of the disease at the top of the canopy. This creates an "unobservable area" in the top-view remote sensing image, manifesting as early systematic missed detections. It is difficult to distinguish between two fundamentally different situations: "there are indeed no lesions at this location" and "the lower leaves at this location are not visible, leading to the failure to observe lesions." Even considering... In cases involving lower leaves, physical intervention using hardware processing is used to expose the lower leaf area for observation or collection, which limits its applicability on a large scale. It relies on external hardware to physically intervene in the canopy structure to achieve local exposure and collection. Due to the limited observability of early lesions, there is a lack of quantification of the effective visibility of lower leaves. It is difficult to form stable and interpretable results for lower leaf disease identification and inference under top-view remote sensing inspection conditions without physical intervention. Therefore, there is a technical problem in the early stage where the low accuracy of disease identification is due to insufficient observability of lower leaves in remote sensing images. Summary of the Invention

[0005] This invention provides a method and system for intelligent identification of maize leaf spot based on remote sensing image analysis. This solves the problem in the prior art where the accuracy of disease identification is low in the early stages due to insufficient observation of lower leaves in remote sensing images. It reduces the chance of early missed detections and improves the feasibility of the results.

[0006] To achieve the above-mentioned objectives, the present invention provides the following technical solution: On the one hand, a smart identification method for maize leaf spot disease based on remote sensing image analysis is provided. This method includes: generating a set of flight routes based on the boundary information of the field to be monitored. The set of flight routes is used to discover and locate early lesions on the lower leaves. Remote sensing images of the area to be monitored are acquired, and the remote sensing images are analyzed to identify suspected lesions. A supplementary flight plan is generated for the identified suspected lesions. At the same time, the severity index of the leaf area and the confidence level of the lower leaves are calculated based on the supplementary flight images to generate the disease distribution results for the entire field. The severity index is used to quantitatively characterize the degree of disease and coverage of the confirmed suspected disease areas. The confidence level of the lower leaves is used to indicate the effective visibility of the lower leaf area in the current image. Based on the disease distribution results, risk level labels are output for the field, the impact assessment of the disease coverage is conducted, and the next round of inspection suggestions are output to the application terminal for display and management based on the impact assessment results.

[0007] On the other hand, a smart identification system for maize leaf spot disease based on remote sensing image analysis is provided. This system includes: a module for identifying suspected locations, a module for generating disease distribution results, and a module for analyzing disease distribution. The module for identifying suspected locations generates a set of flight routes based on the boundary information of the fields to be monitored, acquires remote sensing images of the areas to be monitored, and analyzes the remote sensing images to identify suspected locations. The module for generating disease distribution results generates a supplementary flight plan for the identified suspected locations, and calculates the severity index of the leaf area and the confidence level of the lower leaf based on the supplementary flight images to generate disease distribution results for the entire field. The module for analyzing disease distribution results outputs risk level labels for the fields based on the disease distribution results, conducts an impact assessment on the disease coverage area, and outputs the next round of inspection suggestions to the application terminal for display and management based on the impact assessment results.

[0008] The above technical solution has at least the following advantages compared with the existing technology: 1. This invention identifies suspected disease sites through the first round of flight path discovery. Based on the identification of suspected sites, it uses supplementary flight quantification and lower leaf confidence constraints. Then, based on the risk level of disease distribution, the contiguous area of ​​the spread, and the output of inspection suggestions, it identifies and analyzes lesions. This transforms the problem of early lower leaf lesions being difficult to observe under top-view remote sensing into an engineering-processable workflow. It can not only complete disease monitoring at the field scale, but also output the distribution of the entire field, the boundary of the contiguous area, and the next round of inspection flight path. This enables end-to-end implementation from discovery to management decision-making, reduces early missed detections, and improves the traceability and executability of the results.

[0009] 2. The first set of flight paths is composed of fence flight paths and grid flight paths. The first batch of suspected points obtained by prioritizing edge collection and the second batch of suspected points obtained by whole-field survey are merged and deduplicated to form candidate suspected points. Then, a layered screening mechanism is formed to confirm the suspected points through cross-checking, realizing the point determination path from coarse to fine. Compared with the existing technology that only uses a single grid top view survey, this invention can simultaneously cover the early detection area at the edge of the field and the risk of missed detection inside the field. Through cross-checking, high-resolution verification images are obtained at low altitude and low speed to verify the consistency and stability of candidate points, thereby reducing the amount of invalid work in subsequent supplementary flights and mapping.

[0010] 3. By identifying suspected locations as the core, key verification areas are constructed and local flight routes are generated to carry out regional screening. At the same time, coverage gaps or areas with insufficient image quality are triggered for supplementary flights to obtain complete supplementary flight images. The severity index is further quantified, and the lower leaf confidence score and severity results are entered into the whole field distribution generation and uncertainty labeling process. Compared with the existing technology, which is difficult to distinguish between "disease-free" and "lower leaf not visible" situations, this invention makes explicit constraints on the sufficiency of evidence through regional supplementary flights and lower leaf confidence scores. Early lower leaf lesions are monitored and analyzed, and evidence is supplemented through supplementary flights. At the same time, the output of the same suspected area is improved from qualitative judgment to comparable and traceable quantitative results.

[0011] 4. Disease distribution is aggregated into suspected disease sets by spatial units, and multi-level risk results are output. When determining disease coverage, directional constraints related to disease spread are introduced for contiguous disease identification, thereby generating a set of contiguous disease areas and their boundaries that better reflect the actual expansion mechanism. Based on this, impact assessments are generated using contiguous disease areas as risk sources, and recommendations for the next round of inspections are output. By transforming the identification results into actionable management strategies and prioritizing areas with weak lower leaf observations, evidence from lower leaves is continuously supplemented in subsequent rounds, and systematic missed detections are suppressed. Attached Figure Description

[0012] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying 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.

[0013] Figure 1 Flowchart of a method for intelligent identification of maize leaf spot based on remote sensing image analysis provided in this application embodiment; Figure 2 A flowchart of the first flight path of the field provided in this application embodiment; Figure 3 A flowchart illustrating the disease identification, re-flying, and subsequent processing provided in this application embodiment; Figure 4 This is a schematic diagram of the structure of an intelligent system for maize leaf spot based on remote sensing image analysis provided in an embodiment of this application; Figure 5 A partial view of a remote sensing image of the interior of a maize canopy provided in an embodiment of this application; Figure 6 This is a partial view of a remote sensing image of the lower part of a corn plant provided in an embodiment of this application. Detailed Implementation

[0014] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. All other embodiments obtained by those skilled in the art based on the described embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0015] Unless otherwise defined, the technical or scientific terms used in this invention shall have the ordinary meaning understood by one of ordinary skill in the art to which this invention pertains. The terms “first,” “second,” and similar terms used in this invention do not indicate any order, quantity, or importance, but are merely used to distinguish different components. Similarly, the terms “an,” “a,” or “the,” and similar terms do not indicate a quantity limitation, but rather indicate the presence of at least one. The terms “comprising,” “including,” or “including,” and similar terms mean that the element or object preceding the word encompasses the element or object listed following the word and its equivalents, without excluding other elements or objects. The terms “connected,” “linked,” or “connected,” and similar terms are not limited to physical or mechanical connections, but can include electrical connections, whether direct or indirect.

[0016] It should be noted that the terms "up", "down", "left", "right", "front", and "back" used in this invention are only used to indicate relative positional relationships. When the absolute position of the object being described changes, the relative positional relationship may also change accordingly.

[0017] Embodiment 1 of this invention provides an intelligent identification method for maize leaf spot disease based on remote sensing image analysis, such as... Figure 1The flowchart shown is for an intelligent identification method for maize leaf spot disease based on remote sensing image analysis. This method includes the following steps: S1, generating a set of flight paths based on the boundary information of the field to be monitored. This set of flight paths is used to detect and locate early-stage lesions on lower leaves. Remote sensing images of the monitored area are acquired, and the images are analyzed to determine suspected locations. Lower leaves refer to leaves in the maize plant canopy that are at a lower height relative to the ground surface, attached to several nodes on the lower part of the plant stem. A node is the position formed by the connection between the leaf and the stem, numbered sequentially from the ground surface upwards. Its height changes with plant growth but maintains a stable relative order within the same growth stage. The boundary information of the field to be monitored can be obtained by importing existing plot vector data. After obtaining the boundary, a unified spatial reference is used for subsequent flight path planning and image indexing. The field boundaries are projected onto a unified coordinate system to generate an effective working area within the field. Based on the boundary information, an initial set of flight paths is generated, which includes at least two types: fenced flight paths and grid flight paths. Fenced flight paths are closed paths formed by offsetting a preset distance inward along the field boundary. The offset distance is used to get as close as possible to the high-incidence area at the edge while avoiding the risk of crossing the boundary and obstacles, and to improve the lateral visibility of the lower leaves. Grid flight paths are round-trip paths located within the field boundary. Their grid direction is preferably consistent with the crop row direction to form a better inter-row observation window and reduce plant occlusion. At the same time, to ensure the image stitching capability and the stability of rapid screening, both fenced and grid flight paths can be set with flight parameters that meet the overlap constraint, and the take-off direction is determined by the wind direction, such as a take-off direction that prioritizes the tailwind or crosswind, to reduce the impact of attitude fluctuations on the imaging of the lower leaves.

[0018] It should be understood that when acquiring remote sensing images via camera, the ground sampling distance for the remote sensing images is preferably no greater than 2 cm / pixel, more preferably 0.5 to 2 cm / pixel; when used for early lesion identification and emphasizing visual evidence of lower leaves, the ground sampling distance should not exceed 3 cm / pixel; the remote sensing image acquisition equipment further includes a multispectral imaging sensor, containing at least green, red, red-edge, and near-infrared bands, used to enhance the separability of leaf tissue changes and disease stress through vegetation spectral response in scenarios where visual evidence of lower leaves is sparse or lighting is unstable; the multispectral bands can be a combination or equivalent band combination with green light approximately 550 nm, red light approximately 660 nm, red-edge approximately 735 nm, and near-infrared band approximately 790 nm as center wavelengths; the UAV flight platform is preferably equipped with RTK (Real-Time Kinematic, Real-Time Dynamic Differential Positioning) or PPK (Post-Processed Kinematic) technology. Kinematic (post-event dynamic differential positioning) positioning method enables image exterior orientation elements to have centimeter-level to decimeter-level stability, thereby improving the consistency of multi-temporal image alignment and disease evolution tracking within the same plot; to obtain stable orthophoto stitching and avoid loss of inter-row details, forward overlap refers to the proportion of overlapping area on the ground between two consecutive photos taken on the same flight path (same flight direction), preferably 70% to 85%, and lateral overlap refers to the proportion of overlapping area on the ground between photos taken on the left and right flight paths (parallel flight paths), preferably 60% to 80%; in scenarios with repetitive vegetation texture and high requirements for inter-row details, a combination of approximately 80% forward and approximately 70% lateral overlap can be preferred; it is preferable to collect images during periods of relatively stable lighting to ensure that the visible evidence area of ​​the lower leaves has stable extractability in terms of brightness gradient and texture statistics.

[0019] like Figure 2The diagram shows a flowchart of the first-round flight path for the field provided in this application embodiment. In execution, the drone is first controlled to fly along the fenced-off flight path to complete edge-priority data collection. The purpose of the fenced-off flight path is to utilize the spatial pattern that corn in poorly ventilated areas at the field edges is more susceptible to corn leaf spot disease to form the first batch of suspected locations. After the fenced-off data collection is completed, the images from the fenced-off flight path undergo initial screening. Each frame is first checked for quality, and frames that are obviously blurry, overexposed, or distorted are removed to avoid image quality problems caused by motion jitter, backlighting, etc., from entering the judgment. Unified brightness and color normalization is performed on the usable frames to make the leaf color performance under different times and different lighting conditions comparable, reducing the risk of misreporting cloud shadows and exposure differences as lesions from the source. When necessary, camera calibration parameters can be called to perform distortion correction to ensure the stability of the candidate area morphology statistics. The output then identifies anomalous regions with continuous anomalous bands or patches. Continuous anomalous bands represent lesion clusters extending along the edges, while anomalous patches represent localized clusters of early infection. When the positional proximity and morphological similarity of an anomalous region remain above a threshold across several consecutive frames, the anomaly is deemed stable and marked as the first batch of suspected sites. Positional proximity measures the spatial consistency of the anomalous region across consecutive frames and is calculated using the centroid distance difference of the anomalous region. For each frame, the area and aspect ratio of the anomalous region are extracted to form a morphological feature vector, and the difference between the morphological feature vectors of adjacent frames is calculated to obtain the morphological similarity, which measures the consistency of the anomalous region across consecutive frames. If the anomalous region appears only in a single frame or jumps significantly with each frame, it is not marked as the first batch of suspected sites. The spatial coordinates of the first batch of suspected sites are determined by image metadata or flight control trajectory information.

[0020] After completing the fenced flight path, the drone is controlled to fly along the grid route to complete the field survey and data collection. First, the inter-row observation window is analyzed. The inter-row observation window represents the visible gap between two adjacent rows of crops that is not completely obscured by leaves. If an inter-row observation window exists, priority is given to generating inter-row strip scans along the row direction, and the camera is allowed to observe at a set angle to increase the exposure ratio of lower leaves, thereby improving the visibility and distinguishability of early lesions on lower leaves. The set angle is preset during the deployment stage based on the crop canopy shading mechanism and aerial survey geometric constraints, typically between 15° and 25°. Intervals can increase the perspective depth between rows, while controlling the geometric distortion and secondary occlusion effects introduced by the tilted view within a correctable range. This correctable range is determined during the deployment phase through trial sampling and evaluation of samples at different tilt angles, thus ensuring the stability of subsequent orthorectification and target area extraction. If there are no inter-row observation windows or crop canopy closure leading to insufficient side-view gains, regular grid coverage is used to ensure uniform spatial sampling. After grid acquisition, the grid flight path images undergo initial screening processing consistent with fenced-off methods to output a second batch of suspected locations. The determination of the second batch of suspected locations is based on the images within the crop area. Anomalies are extracted as candidates, and after initial screening, representative points of confirmed candidate regions are output as the second batch of suspected points. The second batch of suspected points is then merged with the first batch to remove duplicates, resulting in candidate suspected points. Deduplication can be based on spatial clustering radius; for example, suspected points with a distance less than a set spatial clustering radius are merged into the same candidate point, while retaining image indices with higher anomaly scores or better imaging quality to reduce redundant verification and re-flying. The set spatial clustering radius is represented by the sum and average of historical spatial clustering radii from the previous suspected point determination process in the database. Anomaly scores are used for quantification. The credibility of lesions in candidate abnormal regions is obtained as follows: After completing brightness and color normalization, the area and aspect ratio of the candidate abnormal region are compared with the preset reasonable lesion range. The degree to which each feature falls within the reasonable range is mapped according to the preset segmented mapping rules to obtain the corresponding conformity score. The average of the conformity scores is then calculated to obtain the abnormal score. The larger the abnormal score, the more likely the candidate region is to be a real lesion. The preset reasonable lesion range represents the range corresponding to the maximum and minimum values ​​of historical reasonable lesions in the historical abnormal score calculation process in the database, including the reasonable maximum and minimum values ​​of lesions.

[0021] The database is a monitoring database for maize leaf spot disease. It is used to uniformly store and manage the data required for disease identification, quantitative assessment, reference value generation, threshold / mapping rule configuration, etc. The database can be deployed on a cloud server. Data sources include UAV first-round inspection and supplementary flight images and corresponding flight records, optional ground close-range sampling and field records, manual annotation / expert review conclusions, and relevant parameters generated during system operation. The original files such as images are stored in the form of objects and their index identifiers are recorded. At the same time, quantitative indicators and rule parameters are saved in structured records and traceability and retrieval are realized through unified primary key association.

[0022] To further reduce false alarms and improve the credibility of suspected sites, a verification process is performed on each candidate site: two short, intersecting straight paths are generated around the site to form a cross-shaped small-area flight path. One short straight path is set along the crop row direction, and the other is set perpendicular to the crop row direction. The purpose of the cross-shaped flight path is to repeatedly observe the same location from two orthogonal directions to distinguish between real lesions and light and shadow, leaf reflection, or partial shading caused by a single viewpoint. During the verification flight, the UAV completes the verification data collection at a set flight altitude and speed. The set flight altitude and speed are both lower than those used during grid flight, thereby obtaining higher resolution verification images and reducing motion blur.

[0023] The verification images undergo suspected point confirmation processing. Candidate points are determined as valid suspected disease points based on the consistency and stability of abnormal morphologies in the two cross directions. When a candidate point simultaneously meets both consistency and stability criteria, it is output as a confirmed suspected point. The consistency criterion is that the abnormal areas observed in the two directions are basically aligned spatially, and their outlines and textures are consistent, thus they can be identified as the same abnormal target. The stability criterion is that the abnormal score corresponding to the anomaly persists along consecutive frames of a short flight segment, and the fluctuation amplitude of the anomaly score does not exceed a preset fluctuation amplitude tolerance. The preset fluctuation amplitude tolerance is represented by the sum and average of historical fluctuation amplitude tolerances in the historical suspected disease point determination process in the database. If the above criteria are not met but the anomaly score is... When a point is found to be within the set abnormal score threshold range or the imaging quality is insufficient to rule out abnormality, the point is output as a point to be further verified. This allows for improved resolution and re-verification during subsequent supplementary flights. The set abnormal score threshold range represents the range corresponding to the maximum and minimum values ​​of the historical abnormal score thresholds in the process of determining historical suspected disease points in the database, including the maximum and minimum values ​​of the abnormal score thresholds. To support the generation of regional supplementary flight plans and the construction of disease distribution results in subsequent steps, the spatial coordinates, abnormal scores, block numbers, and associated image indexes are retained for both suspected and supplementary points. The image index is used to trace the original evidence chain of the point in fence, grid, and cross verification, enabling subsequent supplementary flights to be quickly scheduled with the point as the core and to form verifiable and traceable disease spot discovery and location results throughout the entire process.

[0024] S2 is used to identify suspected locations as the core to generate regional re-flying plans. At the same time, based on the re-flying images, the severity index of the leaf area and the confidence level of the lower leaf are calculated to generate the disease distribution results for the entire field.

[0025] like Figure 3 The diagram shows a flowchart of the disease identification and supplementary flight process provided in this application embodiment. First, suspected locations are identified, key verification areas are constructed, and local investigation flight routes are generated. Next, it is determined whether there are coverage gaps or insufficient image quality. If so, a coverage supplementary flight task is generated and supplementary flight data acquisition is performed; otherwise, existing imagery is used directly. Then, complete supplementary flight images are acquired, and fine-grained identification processing is performed to generate severity indicators and calculate lower leaf confidence scores. Finally, the results are mapped to field spatial units, and the distribution is completed by combining information from adjacent areas to generate the overall field disease distribution results.

[0026] Severity indicators are used to quantify the degree of disease and coverage in confirmed suspected disease areas. Lower leaf confidence indicates the effective visibility of the lower leaf area in the current image. A regional re-flight plan is used to construct a key verification area centered on the identified suspected locations. This key verification area serves as the surrounding area centered on the identified suspected locations. The spatial coordinates of the suspected locations recorded in the first flight path are used as the first-round coordinates. The same abnormal area is then re-detected and located in the re-flight image to obtain the re-location coordinates. The first-round positioning error is the Euclidean distance between the two coordinates in the same geographic coordinate system. When the same location corresponds to multiple frames of observation during re-flight, it is preferable to perform weighted fusion of the re-location coordinates from multiple frames before calculating the Euclidean distance. The first-round positioning error is compared with a set error threshold. When the location positioning error exceeds the set error, the surrounding area is appropriately increased to cover the potential contiguous expansion area around the location. The set error threshold is determined through historical first-round positioning errors in the database. The result is represented by the summation and averaging of the error thresholds; when the positioning is not greater than the set error threshold, the surrounding range can be converged accordingly to improve the efficiency of supplementary flights, thereby ensuring that continuous data is obtained in the key inspection area and that the coverage and degree can be stably quantified. Local flight paths are generated in the key inspection area to perform regional investigation. Regional investigation obtains spatially continuous isometric evidence in the key inspection area to obtain the details of the lesion distribution in the two-dimensional range around the point, so as to statistically analyze the disease coverage boundary, diffusion direction and local coverage. Therefore, regional investigation requires local flight paths to form repeatable and splicable coverage trajectories in the key inspection area, so that the same area can be observed in multiple adjacent flight segments, thereby significantly improving the reliability of subsequent severity quantification and spatial mapping; local flight paths can adopt small-range round trips to form local dense coverage, or they can adopt strip scanning flight paths along the planting row direction to utilize the row channel to improve the observation opportunities of the lower leaves.

[0027] Example A: In a scenario where a suspected disease site is identified internally, the initial imagery reveals multiple scattered anomalies around the site with unclear boundaries, making it difficult to determine whether the disease has progressed from a localized outbreak to a patchy spread. In this case, the area surrounding the site is designated as a key inspection area, and short-range round-trip flights are used to supplement the area. By increasing the coverage density and overlap of these supplementary flight routes, uniform and continuous coverage images are obtained for the key inspection area. This allows for stable statistical analysis of lesion area within the leaf region and identification of spread boundaries, generating severity results that can be directly used for local mapping.

[0028] Example B: For scenarios where there are suspected disease sites in the upper canopy, multiple strip-shaped scanning routes are generated along the planting row direction with the site as the center. During the flight, the camera's line of sight is tilted at a set angle to collect images of the lower part of the plant along the inter-row passage by setting the tilt angle. The supplementary images are mainly used to improve the visibility of the lower leaves and update the confidence of the lower leaves, thereby determining whether the disease has invaded the lower leaf layer and providing a basis for whether to trigger additional shooting or to carry out risk labeling.

[0029] Simultaneously, coverage gaps or insufficient image quality are identified: the coverage area of ​​each image frame is mapped to the field coordinate system and represented as the coverage area. When the ratio of the length of a continuous uncovered gap along the local flight path to the planned coverage length exceeds a set coverage ratio threshold, it is identified as a coverage gap. The set coverage ratio threshold is represented by the sum and average of historical coverage ratio thresholds from the historical coverage assessment process in the database. The impact of each image frame's sharpness meeting a set sharpness threshold is determined as a usable frame, and the ratio of the usable frame's coverage area to the area of ​​the key inspection area is calculated as the effective coverage area percentage. When the effective coverage percentage is low... When a coverage ratio threshold is set, insufficient image quality is identified. The set sharpness threshold is represented by the sum and average of historical sharpness thresholds from the historical image quality assessment process in the database. For areas with coverage gaps or insufficient image quality, a coverage supplementation flight task is generated to fill in missing images and ensure the continuity of the whole field map and the consistency of the background baseline. After acquiring the supplementation flight images, fine recognition processing is performed on them. Fine recognition processing includes at least crop leaf region extraction and non-target region removal to separate background such as soil, shadows, and weeds from the analysis object. The core of fine recognition processing is to separate the analysis object from the whole field map. The images are converged into a reliable crop leaf region to avoid interference from background changes such as soil, shadows, and weeds on lesion identification. Specifically, a leaf region mask is obtained through vegetation segmentation, and non-target areas outside the mask are removed. At the same time, areas with strong shadows and high reflectivity are marked for validity, so that lesion identification and quantification mainly occur on the leaf surface and maintain a more consistent statistical caliber under different lighting and viewing angles. Based on this, lesion identification and quantification are performed within the leaf region to obtain the lesion area ratio, lesion number density, and lesion length density. First, the effective visible area of ​​the leaf is segmented from the supplementary flight image, and then within this area... The diseased area is obtained by segmenting the leaf surface. Simultaneously, leaf pixels are acquired and converted into the effective visible area of ​​the leaf. Diseased area area is obtained by segmenting and counting the number of diseased pixels. The ratio of the diseased area area to the effective visible area of ​​the leaf is used to calculate the diseased area percentage. Within the effective visible area of ​​the leaf, the number of diseased spots is obtained by counting connected components. The diseased number density is then calculated by the ratio of the number of diseased spots to the effective visible area of ​​the leaf. For strip-shaped or linear diseased areas, the centerline is extracted and its length is accumulated to obtain the total diseased length. The diseased length density is then calculated by the ratio of the total length to the effective visible area of ​​the leaf.The ratio of lesion number density to a reference value is used to calculate the number density ratio score, and the ratio of lesion length density to a reference value is used to calculate the length density ratio score. The lesion number density reference value is represented by the sum and average of historical lesion number densities calculated from historical severity indicators in the database. The length density reference value is represented by the sum and average of historical length number densities calculated from historical severity indicators in the database. The severity index is then weighted and fused using preset weights. These preset weights are determined by professionals in the field based on historical baselines, and the sum of the preset weights is 1. The severity index is used to quantitatively characterize the severity and coverage of suspected disease areas, providing quantitative results on the anomalies and coverage of the same suspected area. Simultaneously, the severity index is associated with its corresponding spatial location to form point-level or local area-level identification results, thus providing spatially mappable data input for generating subsequent field-wide disease distribution results.

[0030] Furthermore, for each identified suspected location or its local area, the confidence level of the lower leaf is calculated based on the shooting geometry and occlusion features of the supplementary flight image. First, the flight altitude, camera pitch angle, and the relative relationship between the flight direction and the crop row direction are obtained from the supplementary flight image. Within the effective visible area of ​​the leaf, the area located on the visible side between rows and exhibiting the texture and edge features of the lower leaves of the plant is extracted as the visible evidence area of ​​the lower leaf. Specifically, the visible side between rows of each gap is determined according to the relative orientation of the camera's line of sight and the row direction. The texture and edge features belonging to the lower leaves of the plant are further filtered on the visible side between rows. The acquired remote sensing image is scanned point by point in the inter-row area using a sliding window method. Each window corresponds to a local unit to be judged. For each local unit... The algorithm synthesizes local units that meet a preset texture unit threshold into texture candidate units. The preset texture unit threshold is represented by the average of historical texture unit thresholds determined during the historical suspected point determination process in the database. An edge response algorithm based on brightness gradient is used to calculate the edge intensity index of each pixel in the remote sensing image. Pixels that meet a preset edge intensity threshold are synthesized into edge candidate pixels. The preset edge intensity threshold is represented by the average of historical edge intensity thresholds determined during the historical suspected point determination process in the database. Only when a local region simultaneously contains texture candidate units and is surrounded by edge candidate pixels is the local region confirmed as the lower leaf visible evidence region, and the visibility of the lower leaf visible evidence region on the leaf is statistically analyzed. The proportion of the effective visible area is calculated by determining the pixel brightness within the leaf's effective visible area. Shadow pixels are marked based on a low quantile threshold, and highly reflective pixels are marked based on a high quantile threshold. The proportion of shadow pixels and highly reflective pixels to the total number of pixels in the leaf's effective visible area is then calculated to determine the degree of occlusion and distortion. The proportion of lower leaf visibility evidence and the degree of occlusion and distortion are weighted and fused to obtain the lower leaf confidence score. This lower leaf confidence score represents the effective visibility of the lower leaf region in the current image and provides an explicit constraint on the reliability of the quantification results. When the lower leaf confidence score is lower than a preset threshold, it is considered that there is a risk of insufficient lower leaf observation at that point. Anomaly scores for the canopy visible area at that point are obtained, and the scores are calculated based on the score of that point and its neighboring areas. The similarity of disease among adjacent suspected sites within the range is used as the neighborhood similarity score. Then, the upper canopy anomaly score and the neighborhood similarity score are summed and averaged to obtain the lower leaf risk score. The lower leaf risk score is used as a supplementary component or uncertainty label for the severity index of the site, so as to provide a conservative indication of potential lower leaf risk and clearly mark the source of uncertainty even with insufficient observation. At the same time, an additional supplementary monitoring task is triggered for the site. The confidence of the lower leaf is improved by further reducing the altitude, changing the gimbal pitch angle, or conducting lateral supplementary flights between rows. After obtaining more visual evidence, the inferred results and severity index are updated, thereby reducing early underreporting of lower leaves and improving the verifiability, traceability, and engineering feasibility of the final disease distribution results.

[0031] Using the boundaries of the fields to be monitored as constraints, a whole-field spatial carrying capacity unit is established. A regular grid is generated using a preset grid side length as the spatial resolution, and then trimmed to obtain grid cells. Simultaneously, the point-level or local area-level identification results obtained during the supplementary flight phase are mapped to the corresponding grid cells according to their spatial location. Local area-level identification results can be weighted and distributed to each grid cell according to their overlap area with the grid cell. Point-level results can be projected onto neighboring grid cells based on their location or a set influence radius. For each grid cell, its corresponding severity index and lower-leaf confidence are aggregated. Local area-level identification results covering the grid cell are preferentially used as the primary evidence for that grid cell. When only point-level results exist, those falling into the grid cell are considered primary evidence. The results of points within and around a grid are weighted and aggregated to obtain a grid-level severity characterization. The weighting is constrained by the distance from the point to the grid center and the corresponding lower leaf confidence level, so that evidence with sufficient visibility in the lower leaf has a higher weight in the grid aggregation, while the uncertainty labeling of the lower leaf confidence level is not sufficient and is reflected in the grid result as confidence attenuation or risk supplementary components. After obtaining the grid-level severity characterization and corresponding confidence labeling of each grid unit in the whole field, it is output as the disease distribution result of the whole field according to the preset mapping rules, thereby realizing the unified mapping from the point-level and local area-level quantitative results of the supplementary imagery to the continuous and interpretable disease distribution results of the whole field.

[0032] S3 outputs risk level labels for fields based on disease distribution results, conducts an impact assessment on disease coverage, and outputs the next round of inspection suggestions to the application for display and management based on the impact assessment results. The disease distribution results are structured to form risk level labels, where the disease distribution results are spatially represented by grid cells carrying severity indicators and lower leaf confidence and uncertainty labels. On this basis, grid cells with severity indicators reaching the preset risk threshold are aggregated into a suspected disease candidate set, and a list of confirmed suspected points from the re-flying phase is summarized as point-level evidence. The list of confirmed suspected points includes at least point coordinates, associated image index, severity indicator, and lower leaf confidence. Field risk level labels are output, so that field risk reflects both the most severe local situation and the coverage area. At the same time, the application output includes at least the list of confirmed suspected points, the boundary of contiguous disease areas, and the corresponding severity level information. The severity level can be obtained by interval mapping of severity indicators and bound to color or symbol encoding to support managers to quickly identify and filter on the map.

[0033] When defining the disease coverage area, grid cells exceeding the suspected threshold range set by the disease distribution results are aggregated to form the boundary of contiguous disease areas. However, to avoid false or missed contiguous areas due to geometric adjacency alone, a propagation direction constraint is introduced to improve the rationality of contiguous area determination. The set suspected threshold range is represented by the sum and average of historical suspected threshold ranges obtained from the historical disease coverage area calculation process in the database. When wind direction information is obtained, the wind direction vector is used as the priority propagation direction, and adjacent grid cells located at a preset wind direction propagation distance are given higher weight in contiguous area determination, making them more likely to be merged into the same contiguous area under the same severity conditions. The preset wind direction propagation distance is the sum and average of historical wind direction propagation distances obtained from the historical disease coverage area calculation process in the database. The results show that adjacent grids beyond the preset wind direction propagation distance are assigned relatively low weights, making the contiguous boundaries more consistent with the directional expansion characteristics of the disease under wind-driven propagation. When planting row information is obtained, the row direction is used as the priority direction for strip expansion, and strip-shaped connections between adjacent grids along the row direction are assigned higher weights, making linear or strip-shaped expansions easier to identify as the same contiguous area. Meanwhile, cross-row connections perpendicular to the row direction are assigned lower weights to suppress lateral false connections caused by local noise points. When both wind direction and row direction are available, the weights of the two types of directions are combined, giving the highest priority to downwind and row-oriented connections. This naturally forms contiguous disease area boundaries that conform to the propagation mechanism in spatial aggregation, and outputs them as polygons at the application end to support on-site inspection and operation planning.

[0034] Based on the formed contiguous disease areas and their severity levels, an impact assessment result on other areas of the field is generated. The role of the impact assessment result is to transform the discovered disease status into management prompts for the next expansion areas and areas requiring key attention. Therefore, contiguous disease areas are used as risk sources, and an extended concern zone is constructed by combining wind direction information, row direction information, and information on wet areas or low-lying areas of the field: When wind direction information is available, the extended concern zone is formed by extending downwind along the boundary of the contiguous disease area, and the near-field and far-field concern ranges can be output in layers according to distance to represent short-term and medium-term expansion risks; when wet areas or low-lying areas are available, the wet area zone is regarded as an environmentally sensitive area where the disease is more likely to continue to develop. Wet areas adjacent to or overlapping with contiguous disease areas downwind are given priority to be marked as high concern areas, so that the impact assessment reflects not only the direction of spread but also the environmental suitability; at the same time, based on the confidence distribution of the lower leaves within the field, areas with low long-term confidence are marked as weak observation areas. When the weak observation area overlaps with the extended concern zone, its concern level is given priority to be increased to avoid systematic missed detections due to shading.

[0035] Based on the above impact assessment results, recommendations for the next round of inspections are generated and sent to the application for display and management. These recommendations include at least priority inspection areas, recommended inspection periods, and recommended inspection routes. Priority inspection areas are determined by combining the results of the contiguous disease area itself, its downwind expansion zone of concern, the high-concern area in the wet zone, and the weak observation area of ​​the lower leaves, ensuring that inspection resources prioritize coverage of the areas most likely to expand and most easily missed. The recommended route framework for the next round of inspections includes a circular route along the boundary of the contiguous disease area for monitoring boundary expansion, a strip-shaped route along the downwind expansion zone of concern for tracking the direction of propagation, and routes targeting the lower leaves. Interrow strip scanning of areas with persistently low confidence levels and high risk scores in lower leaves is used to form a complete observation of the lower leaves. This allows the next round of inspections to both monitor changes in the extent of the disease and improve the sufficiency of evidence for the lower leaves. At the same time, the inspection recommendations are displayed in a structured manner on the application side, allowing managers to view the triggering reasons for each priority inspection area (such as the severity level of contiguous areas, downwind expansion weight, wet area superposition, or insufficient confidence level of lower leaves). The flight path can be fine-tuned according to actual operational constraints before being issued for execution. This enables continuous monitoring and rolling investigation driven by disease distribution results, thereby reducing missed detections, improving detection efficiency, and reducing invalid flights.

[0036] In Embodiment 2 of this invention, based on Embodiment 1, when the confidence level of the lower leaf is insufficient, inference assessment is no longer used as the main compensation method. Instead, visual evidence of the lower leaf is obtained through repeatable and convergent dedicated shooting geometry. After sufficient evidence is obtained from multiple frames, the severity index is directly updated, thereby improving the verifiability and traceability of the lower leaf level and reducing the spread of uncertainty caused by inference. Specifically, after identifying suspected locations as the core to construct key verification areas and generate local flight paths, unlike Embodiment 1, the calculation method of the lower leaf confidence level is replaced by the proportion of lower leaf visual evidence. That is, after extracting the leaf region from the re-shot images, the visible leaves are further divided into lower leaf candidate visible regions according to height or morphology layers within the leaf region. The proportion of effective pixels identified as lower leaf candidate visible regions and their cross-frame consistency in the multi-view sequence are statistically analyzed to obtain the lower leaf confidence level, which is used to characterize whether the lower leaf has been sufficiently observed at the current location. When the lower leaf confidence level is lower than a preset threshold, this embodiment does not directly enter the lower leaf lesion inference assessment, but instead uses the lower leaf lesion inference assessment as the primary compensation method. Insufficient confidence is used as a criterion for non-convergence of supplementary imaging, triggering additional supplementary imaging subtasks. These supplementary imaging subtasks increase the probability of lower leaf exposure by further reducing the height, shortening the lateral approach distance, or increasing the field of view of the fan-shaped scan. The confidence of the lower leaf is recalculated after each round of supplementary imaging until it reaches a threshold, thus forming a complete, terminateable, and convergent process for acquiring lower leaf evidence. After the lower leaf confidence reaches the threshold, lesion identification and quantification are performed within the candidate visible area of ​​the lower leaf, and the amount of lower leaf lesions from different perspectives is recorded. The results are aggregated in multiple frames based on consistency and clarity to obtain stable estimates of lesion area ratio, lesion number density, or lesion length density at the lower leaf level. These estimates are then combined with the quantitative results from the upper canopy or the entire leaf area to update the severity index, providing direct observational evidence to support the severity index at the lower leaf level. In this embodiment, the approach for lower leaf-related processes has been changed from primarily relying on extrapolation and compensation when confidence is insufficient to primarily relying on supplementary evidence obtained through multi-view re-capture when confidence is insufficient. This significantly improves the visibility of early lesions on lower leaves.

[0037] like Figure 4 The diagram shown is a structural schematic of an intelligent system for maize leaf spot based on remote sensing image analysis provided in this application embodiment, including: a module for identifying suspected locations, a module for generating disease distribution results, and a disease distribution analysis module.

[0038] The suspected location identification module is used to generate a set of flight routes based on the boundary information of the fields to be monitored, acquire remote sensing images of the area to be monitored, analyze the remote sensing images to identify suspected locations, and then transmit the identified suspected locations and the quality information of the first round of inspection images to the disease distribution result generation module for subsequent supplementary flights and quantification.

[0039] The module for generating disease distribution results is used to identify suspected locations and generate regional re-flying plans. At the same time, it calculates the severity index of the leaf area and the confidence level of the lower leaf based on the re-flying images to generate disease distribution results for the entire field. This module transmits the disease distribution results for the entire field, the re-flying tasks and quantitative records to the disease distribution analysis module for risk assessment and inspection recommendations.

[0040] The disease distribution analysis module is used to output risk level labels for fields based on disease distribution results, conduct impact assessments on disease coverage, and output the next round of inspection suggestions to the application for display and management based on the impact assessment results. This module obtains risk level labels, key points of the impact assessment report, and the next round of inspection suggestions, and pushes them to the application.

[0041] The modules interact sequentially through data interfaces. The module for identifying suspected locations is connected to the module for generating disease distribution results, and the module for generating disease distribution results is connected to the module for analyzing disease distribution. The modules for identifying suspected locations, generating disease distribution results, and analyzing disease distribution are all connected to the task and data database for reading and writing task configurations, image indexes, location results, re-flying tasks, and analysis results of maize leaf spot disease on lower leaves.

[0042] like Figure 5 The image shown is a partial view of a remote sensing image of the interior of the maize canopy provided in this application embodiment. The leaves are arranged relatively densely, still maintaining a high degree of green coverage. Multiple overlapping leaves are visible. The subject of the image is mainly located in the middle and upper leaf layer of the plant. Several marked areas in the image are lesions of maize leaf spot disease. Their common characteristics are: the lesions are strip-shaped or spindle-shaped structures extending along the leaf veins, and the color transitions from light green to light yellow or light brown. The boundaries of the lesions are relatively clear, forming a significant contrast with the surrounding healthy leaf tissue. The lesions are of limited length and are discretely distributed. From the perspective of remote sensing features, this image shows abnormal linear textures formed by the lesions on the leaf surface. The leaf surface reflectance is locally reduced in the lesion area, but the leaf as a whole still maintains good integrity and tension.

[0043] like Figure 6 The image shown is a partial view of a remote sensing image of the lower part of a maize plant provided in this application embodiment. The maize plant leaves were also photographed. It can be observed that the lower leaves are mature and significantly shaded. Analyzing the image from the perspective of composition and spatial hierarchy: the lesions are mostly concentrated in the areas of longer and more mature leaves. The lesion stripes are longer and more continuous, and some show a merging trend. The marked lesions show continuous light brown stripes extending along the leaf veins. The severity of the lesions is significantly greater than that of the leaves inside the canopy. There are multiple parallel lesions on the same leaf, and the overall color of the leaf is darker, reflecting the physiological state closer to the lower part of the plant. These characteristics are consistent with the typical transmission pattern of maize leaf spot during plant development. The disease often occurs first in the lower leaves and spreads upward under suitable conditions.

[0044] The above are merely specific embodiments of the present invention, but the scope of protection of the present invention is not limited thereto. The scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A method for intelligent identification of maize leaf spot disease based on remote sensing image analysis, characterized in that, Includes the following steps: S1. A set of flight routes is generated based on the boundary information of the field to be monitored. The set of flight routes is used to detect and locate early lesions on the lower leaves, acquire remote sensing images of the area to be monitored, and analyze the remote sensing images to determine suspected locations. S2, to determine the suspected locations as the core to generate a regional supplementary flight plan, and at the same time calculate the severity index of the leaf area and the confidence of the lower leaf to generate the disease distribution results of the whole field. The severity index is used to quantitatively characterize the disease degree and coverage of the confirmed suspected disease area, and the lower leaf confidence is used to indicate the effective visibility of the lower leaf area of ​​the current image. S3 outputs risk level labels for fields based on disease distribution results, conducts impact assessments on disease coverage, and outputs the next round of inspection suggestions to the application for display and management based on the impact assessment results.

2. The intelligent identification method for maize leaf spot disease based on remote sensing image analysis according to claim 1, characterized in that, The generated flight route set includes: Obtain the boundary information of the fields to be monitored and divide the fields into several monitoring blocks; The first set of flight routes is generated based on boundary information. The first set of flight routes includes fenced flight routes at the edge of the plot and grid flight routes covering the entire field. The fenced flight routes are used to monitor areas with poor ventilation at the edge of the plot that are more susceptible to corn leaf blight. The grid flight routes are used to monitor diseases inside the plot.

3. The intelligent identification method for maize leaf spot disease based on remote sensing image analysis according to claim 1, characterized in that, The analysis of remote sensing images to determine suspected locations includes executing flight routes to identify candidate suspected locations: Control the drone to fly along the fence route to complete edge-priority data collection, determine the location of suspected defects based on continuous abnormal bands or abnormal patches in the images, and analyze and record the first batch of suspected points; Then, control the drone to fly along the grid route to complete the field survey and collect data to output the second batch of suspected locations; The second batch of suspected locations was merged with the first batch of suspected locations to remove duplicates, resulting in candidate suspected locations.

4. The intelligent identification method for maize leaf spot disease based on remote sensing image analysis according to claim 3, characterized in that, The process of completing the field survey by flying along a grid route includes determining the grid flight method based on the observation windows between the rows of the field: If an inter-row observation window exists, priority is given to generating inter-row strip scans along the row direction, and the camera is allowed to observe at a set angle to increase the proportion of lower leaves exposed, thereby improving the visibility and distinguishability of early lesions on lower leaves. If there is no inter-row observation window or the crop canopy is closed, resulting in insufficient side-view benefits, a regular grid cover should be used to ensure spatially uniform sampling.

5. The intelligent identification method for maize leaf spot disease based on remote sensing image analysis according to claim 3, characterized in that, The process of obtaining candidate suspected locations includes a review process for each candidate location: A cross-shaped flight path is generated, consisting of two intersecting straight-line paths with a set distance, centered on the candidate suspected location. One short straight-line path is set along the crop row direction, and the other short straight-line path is set perpendicular to the crop row direction. Control the drone to complete the verification data collection at a flight altitude and speed lower than the grid line along the confirmed route, so as to obtain high-resolution verification images of the candidate point; The verification images are processed for point confirmation. Based on the consistency and stability of abnormal morphologies in the two directions of the cross, the locations of suspected defects are determined. If the consistency and stability requirements are met, the candidate suspected location is determined as a confirmed suspected location. If the consistency and stability requirements are not met, the candidate suspected site will be determined as a site requiring further verification.

6. The intelligent identification method for maize leaf spot disease based on remote sensing image analysis according to claim 1, characterized in that, The process involves generating a supplementary flight plan for the identified suspected locations, while simultaneously calculating the severity index of the blade region and the confidence level of the lower blade. To identify suspected locations, a key verification area is constructed, and local flight routes are generated within the key verification area to carry out regional investigation. For areas identified as having coverage gaps or insufficient image quality, a coverage supplement flight mission is generated to fill in the missing images and obtain complete supplement flight images. Fine-grained identification processing is performed on the re-fly images to extract crop leaf regions. Lesion identification and quantification are performed to obtain lesion area ratio, lesion number density, and lesion length density. The severity index is obtained by fusion. The confidence level of the lower leaf is calculated based on the shooting geometry and occlusion features of the re-fly images.

7. The intelligent identification method for maize leaf spot disease based on remote sensing image analysis according to claim 6, characterized in that, The results of the disease distribution across the entire field are as follows: Map the point-level or local area-level disease identification results obtained from supplementary aerial images to the field spatial carrying unit; The severity index of disease and the confidence level of the lower leaf are aggregated for each spatial bearing unit to form a unit-level disease characterization; By combining information on diseases in adjacent areas and constraints on the direction of spread, the results of disease distribution across the field, reflecting the severity, spatial continuity, and uncertainty of the disease, are obtained.

8. The intelligent identification method for maize leaf spot disease based on remote sensing image analysis according to claim 1, characterized in that, The method of outputting risk level labels for fields based on disease distribution results includes contiguous area identification with constraints on the direction of spread: The disease distribution results are aggregated by spatial unit to form a suspected disease set, and multi-level risk results are output for the field based on the disease and observable information of the lower leaves. When defining the extent of disease coverage, directional constraints related to disease propagation in lower leaves are introduced to determine the contiguous nature of spatially adjacent units: When based on wind direction information, higher weight is given to the connectivity of adjacent areas downwind for contiguous identification. When based on planting row direction information, higher weight is given to the strip-like connectivity along the row direction for contiguous identification, so that the contiguous results are more consistent with the directional characteristics of disease spread. Generate a set of contiguous disease areas and their corresponding severity levels, and output the boundary information of the contiguous disease area set.

9. The intelligent identification method for maize leaf spot disease based on remote sensing image analysis according to claim 8, characterized in that, The process of outputting next round inspection recommendations to the application for display and management based on the impact assessment results includes: Based on the contiguous disease areas and their severity levels, an impact assessment of the disease's potential spread to other areas of the field is generated. The impact assessment includes at least indications of the possible direction of disease expansion and suggestions on the scope of attention for surrounding areas. Output the next round of inspection recommendations, including priority inspection areas and their suggested routes to support continuous monitoring and rolling scheduling. The next round of inspection recommendations will prioritize areas where the lower leaf confidence and lower leaf risk scores do not meet the set conditions, so as to form a complete observation of the lower leaf.

10. A system applying the intelligent identification method for maize leaf spot based on remote sensing image analysis as described in any one of claims 1-9, characterized in that, include: The module includes modules for identifying suspected locations, generating disease distribution results, and analyzing disease distribution. The suspected location determination module is used to generate a set of flight routes based on the boundary information of the field to be monitored, acquire remote sensing images of the area to be monitored, and analyze the remote sensing images to determine suspected locations. The module for generating disease distribution results is used to determine suspected locations as the core to generate a supplementary flight plan for the region. At the same time, it calculates the severity index of the leaf region and the confidence level of the lower leaf based on the supplementary flight images to generate disease distribution results for the entire field. The disease distribution analysis module is used to output risk level labels for fields based on disease distribution results, conduct impact assessments on disease coverage, and output the next round of inspection suggestions to the application terminal for display and management based on the impact assessment results.

Images