A method and system for monitoring corn small spot disease in saline-alkali soil by low-altitude remote sensing

Abstract

The present application provides a kind of saline-alkali soil corn small spot low-altitude remote sensing monitoring method and system, it is related to precision agriculture technical field.The saline-alkali soil corn small spot low-altitude remote sensing monitoring method of the present application specifically includes: periodically collecting the multi-source remote sensing image of target corn field, generates the multi-channel orthophoto map covering the target corn field;The weed in the target corn field is carried out instance segmentation, and the disease diagnosis is carried out to the weed instance segmented, based on the geographic coordinates of one or more pathogenic weeds, in combination with historical weather data in preset time window, generates the space-time latent risk map covering the target corn field;The target corn field is divided into multiple regions of different risk levels, and the differentiated detection network and adaptive preference decision function matched with the risk level are used to detect the disease of corn plant in the region.

Description

Technical Field

[0001] This invention relates to the field of precision agriculture technology, and more specifically, to a crop disease monitoring method using low-altitude remote sensing by unmanned aerial vehicles and artificial intelligence image recognition technology, which is particularly suitable for early warning and control of maize leaf spot disease under saline-alkali stress. Background Technology

[0002] As a major global food, feed, and industrial raw material crop, the safe production of maize plays a crucial role in ensuring food security and promoting economic development. However, maize is frequently threatened by various diseases during its growth cycle. Among them, maize leaf spot (also known as southern leaf blight), caused by a fungus, is one of the key diseases affecting maize yield and quality. This disease is highly prevalent in high-temperature and high-humidity environments, spreads rapidly, and can cause widespread leaf necrosis in a short period, severely affecting the plant's photosynthesis and leading to a significant drop in yield. Therefore, timely and accurate monitoring and early warning of maize leaf spot are essential prerequisites for effectively controlling its spread and reducing economic losses.

[0003] With increasingly scarce land resources, large areas of saline-alkali land are being developed for agricultural production, posing a more severe challenge to maize cultivation. Saline-alkali soil environments exert continuous abiotic stress on maize plants, leading to inhibited growth, decreased root vitality, and physiological symptoms such as yellowing, wilting, and even leaf edge scorching. The tricky part is that these physiological traits caused by saline-alkali stress are morphologically very similar to the early symptoms of maize leaf spot disease, such as chlorosis and pale yellow spots, making them easily confused visually. This phenomenon of similar symptoms with different origins significantly interferes with accurate disease diagnosis. Traditional monitoring methods struggle to isolate environmental stress factors, often leading to misdiagnosis (mistaking saline-alkali stress for disease) or underdiagnosis (mistaking early-stage disease for saline-alkali stress), not only delaying optimal control but also potentially increasing costs and polluting the environment due to pesticide overuse.

[0004] Currently, monitoring methods for crop diseases are mainly divided into two categories: traditional manual inspection and modern remote sensing technology. Traditional manual field inspection methods rely entirely on the experience and visual observation of plant protection personnel. When facing large-scale planting scenarios, the shortcomings of this method are obvious: it is inefficient, time-consuming and labor-intensive, and cannot achieve rapid and comprehensive coverage of vast fields; secondly, the diagnostic results are highly subjective and inconsistent, easily affected by personal experience, sense of responsibility and weather conditions, resulting in poor reliability and consistency of monitoring results; thirdly, detection is delayed, and by the time plant protection personnel find obvious diseased plants in the field, the disease has often already occurred and spread for some time, missing the golden window for prevention and control at the lowest cost.

[0005] To overcome the shortcomings of manual inspections, technologies utilizing low-altitude remote sensing platforms such as drones equipped with sensors for crop monitoring have emerged. Existing remote sensing monitoring methods typically employ a comprehensive scanning and brute-force identification strategy, attempting to search for lesions on every single corn leaf in aerial images using image recognition algorithms. However, this approach faces new bottlenecks in practical applications. On one hand, corn planting density is high and the number of leaves is extremely large; high-precision segmentation and identification of all leaves requires enormous computational resources, resulting in slow processing speeds and failing to meet the near-real-time early warning needs of large-scale fields. On the other hand, most existing identification models ignore the epidemiological patterns of disease spread in the field and lack intelligent search strategies. They treat the entire field as a homogeneous monitoring object, failing to effectively utilize the key characteristic that diseases typically erupt from specific infection sources in a point-like manner before spreading outwards. More importantly, in complex compound stress environments such as saline-alkali land, existing algorithm models often suffer from limited training data and are unable to cope with the complex situation of physiological lesions and fungal diseases intertwined, significantly reducing their robustness and practical application effectiveness.

[0006] In conclusion, developing an intelligent monitoring method that can adapt to the complex environment of saline-alkali land, conform to the laws of disease transmission, and balance efficiency and accuracy has become an urgent technical challenge to be solved in the field of precision plant protection. Summary of the Invention

[0007] This embodiment provides a low-altitude remote sensing monitoring method for maize leaf spot disease in saline-alkali land, which includes the following steps: periodically acquiring multi-source remote sensing images of the target maize field, and preprocessing the multi-source remote sensing images to generate a multi-channel orthophoto map covering the target maize field;

[0008] Based on the multi-channel orthophoto image, weeds in the target cornfield are segmented into instances, and disease diagnosis is performed on the segmented weed instances to identify one or more pathogenic weeds.

[0009] Based on the geographical coordinates of the one or more pathogenic weeds, and combined with historical meteorological data within a preset time window, a spatiotemporal latent risk map covering the target cornfield is generated.

[0010] Based on the spatiotemporal latent risk map, the target cornfield is divided into multiple areas with different risk levels, and a differentiated detection network and adaptive preference decision function that match the risk level are used to detect diseases in the corn plants within the area.

[0011] Generate comprehensive monitoring and risk warning maps or variable spray prescription maps that include maize disease diagnosis results.

[0012] The multi-source remote sensing images are preprocessed using a two-level feature matching strategy, which constructs a feature matching matrix from the original images. any point in To its precise position in the coordinate system of the final stitched image mapping function Its definition is:

[0013]

[0014] in, The global affine transformation matrix, the It is calculated by matching only stable feature points such as soil surface between adjacent images; For non-rigid compensation vectors, the Through the The dense optical flow is calculated for the pixels in the vegetation area within the overlapping region of the initially aligned images.

[0015] Weeds are segmented into instances using a progressively optimized segmentation network, which includes an optimization module that iteratively optimizes the mask boundaries in a cascaded manner.

[0016] In the Each iteration stage leads to the final refined mask. From the previous stage mask And the currently predicted mask update graph It is composed of combinations, and the combination method is as follows:

[0017]

[0018] in, and The first Stage and First Refined masking for each stage; It is in the The phase specifically targets the high-resolution mask update map predicted for the boundary region; This is a bilinear interpolation upsampling operation. Pixel-level multiplication; To use the mask from the previous stage The calculated boundary region mask is defined as the mask The Euclidean distance between the center and the contour pixel set is less than or equal to a preset distance threshold. A set of pixels.

[0019] The progressive optimization segmentation network includes a context guide head and an instance masking head; the context guide head is used to generate a global semantic context graph that distinguishes corn, weeds, and background / soil; the instance masking head is used to generate an initial coarse mask for each candidate region;

[0020] Before iterative optimization, a context instance feature fusion module is also included, which is used to fuse the instance features of the instance masking tool. Contextual features of the context header Perform fusion to generate a fused feature map. .

[0021] Disease diagnosis was performed on the isolated weed instances using a dual-stream heterogeneous feature fusion diagnostic network.

[0022] The diagnostic network includes a visual network and a multispectral network, and the feature vectors output by the two networks are fused through a channel attention fusion module. and The feature vectors are then fused to obtain the final fused feature vector. The calculation method is as follows:

[0023] in, and These are the feature vectors output by the visual network and the multispectral network, respectively. This indicates splicing along the channel dimension; This is a channel-dimensional attention weight vector.

[0024] Construct a time-series infection efficiency weight function Used to quantify historical moments Under specific meteorological conditions, the overall efficiency of spore release from release to successful infection is defined as:

[0025] in, and They are time points Hourly average relative humidity and hourly cumulative rainfall; It is a humidity effect function used to describe the effect of humidity on spore activity; It is a rainfall effect function used to describe the influence of rainfall on spore deposition and germination.

[0026] Generated Spatiotemporal Latent Risk Map any one of its grid cells Total latent risk value In order to be within the preset epidemiological time window Inside, received from all The weight of the pathogenic weeds and their time-series infection efficiency. The weighted sum of spore concentrations:

[0027] in, For at any time By the Disease-causing weeds The concentration of spores produced; Global coordinates In terms of pathogens Using the origin as the time The wind direction is the relative coordinate in the local coordinate system of the x-axis.

[0028] The adaptive preference decision function used includes an adaptive decision threshold function. The output value of this function is related to the current detected point. Latent risk value Related, which is defined as:

[0029]

[0030] Final diagnostic category :

[0031]

[0032] in, As a baseline threshold, As a preference factor; To detect network peers The original predicted probability of the presence of lesions.

[0033] The generated variable spray prescription map contains grid cells for each pesticide application. Calculated recommended dosage :

[0034] in, and These are the set minimum and maximum effective dosages, respectively. This is the overall severity index for that grid cell, which is determined by the latent risk value at that point. and the pixel density of the lesion in its local neighborhood We obtain the weighted sum.

[0035] This embodiment also provides a low-altitude remote sensing monitoring system for maize leaf spot disease in saline-alkali land, the system comprising:

[0036] Data acquisition module: Periodically acquires multi-source remote sensing images of the target cornfield, and preprocesses the multi-source remote sensing images to generate a multi-channel orthophoto map covering the target cornfield;

[0037] Weed pathogen identification module: Based on the multi-channel orthophoto image, weeds in the target cornfield are segmented into instances, and disease diagnosis is performed on the segmented weed instances to identify one or more pathogenic weeds;

[0038] Spatiotemporal latent risk map generation module: Based on the geographical coordinates of the one or more pathogenic weeds, combined with historical meteorological data within a preset time window, a spatiotemporal latent risk map covering the target cornfield is generated;

[0039] Corn leaf spot disease detection module: Based on the spatiotemporal latent risk map, the target cornfield is divided into multiple areas with different risk levels, and a differentiated detection network and adaptive preference decision function that match the risk level are used to detect the corn plants in the area.

[0040] Monitoring module: Generates comprehensive monitoring and risk warning maps or variable spray prescription maps that include corn disease diagnosis results.

[0041] An electronic device includes a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the aforementioned low-altitude remote sensing monitoring method for maize leaf spot disease in saline-alkali land.

[0042] A computer-readable storage medium storing a computer program that, when executed by a processor, implements the aforementioned low-altitude remote sensing monitoring method for maize leaf spot disease in saline-alkali land.

[0043] In the data preprocessing stage, this invention employs a two-level feature matching strategy to address the difficulties in image stitching caused by the non-rigid movement of vegetation due to factors such as wind. This strategy utilizes the stable soil surface for global rigid alignment, and then uses dense optical flow to compensate for the non-rigid deformation of the dynamic vegetation canopy, ensuring that the final multi-channel orthophoto map has extremely high spatial accuracy and internal consistency.

[0044] This invention realizes a proactive early warning technology path that first identifies the source of infection (weeds) and then infers the risk. To achieve this goal, this invention designs a progressively optimized segmentation network for weed control in cornfields. Through context-guided and region-focused optimization, it solves the problems of weed targets being submerged in the corn background and the blurred boundaries caused by occlusion. Furthermore, when diagnosing whether weeds are diseased, a dual-stream heterogeneous feature fusion diagnostic network is used, which can simultaneously analyze the visible light and multispectral features of the target, enabling it to distinguish between physiological traits of salt stress and pathological symptoms of small leaf spot disease.

[0045] This invention utilizes confirmed pathogenic weeds, combined with historical meteorological data and scientific spore dispersal models, to construct a spatiotemporal latent risk map.

[0046] This invention utilizes a spatiotemporal latent risk map to categorize fields into high, medium, and low-risk zones, and deploys refined and coarse detection networks with varying performance and costs to the corresponding areas. Simultaneously, its adaptive preference decision function dynamically adjusts the aggressiveness or conservatism of the diagnosis based on the risk prior of each detection point, achieving an optimal balance between ensuring high detection rates in high-risk areas and reliability in low-risk areas. This optimizes the allocation of computational resources and automatically transforms all monitoring and analysis results into variable spraying prescription maps that can be directly executed by automated equipment such as agricultural drones, realizing a complete closed loop from data collection and intelligent analysis to precise decision-making and execution. Attached Figure Description

[0047] Figure 1 This is a flowchart for low-altitude remote sensing monitoring of corn leaf spot disease in saline-alkali land. Detailed Implementation

[0048] The embodiments of this application will now be described in detail with reference to the accompanying drawings.

[0049] The following specific examples illustrate the implementation of this application. Those skilled in the art can easily understand other advantages and effects of this application from the content disclosed in this specification. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of them. This application can also be implemented or applied through other different specific embodiments, and the details in this specification can also be modified or changed based on different viewpoints and applications without departing from the spirit of this application. It should be noted that, in the absence of conflict, the following embodiments and features in the embodiments can be combined with each other. Based on the embodiments in this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0050] It should be noted that various aspects of embodiments within the scope of the appended claims are described below. It will be apparent that the aspects described herein can be embodied in a wide variety of forms, and any particular structure and / or function described herein is merely illustrative. Based on this application, those skilled in the art will understand that one aspect described herein can be implemented independently of any other aspect, and two or more of these aspects can be combined in various ways. For example, any number and aspects set forth herein can be used to implement the device and / or practice the method. Additionally, this device and / or method can be implemented using structures and / or functionalities other than one or more of the aspects set forth herein.

[0051] Additionally, specific details are provided in the following description to facilitate a thorough understanding of the examples. However, those skilled in the art will understand that practice can be carried out without these specific details.

[0052] This embodiment provides a low-altitude remote sensing monitoring method for maize leaf spot disease in saline-alkali land, which includes the following steps: periodically acquiring multi-source remote sensing images of the target maize field, and preprocessing the multi-source remote sensing images to generate a multi-channel orthophoto map covering the target maize field;

[0053] Based on the multi-channel orthophoto image, weeds in the target cornfield are segmented into instances, and disease diagnosis is performed on the segmented weed instances to identify one or more pathogenic weeds.

[0054] Based on the geographical coordinates of the one or more pathogenic weeds, and combined with historical meteorological data within a preset time window, a spatiotemporal latent risk map covering the target cornfield is generated.

[0055] Based on the spatiotemporal latent risk map, the target cornfield is divided into multiple areas with different risk levels, and a differentiated detection network and adaptive preference decision function that match the risk level are used to detect diseases in the corn plants within the area.

[0056] Generate comprehensive monitoring and risk warning maps or variable spray prescription maps that include maize disease diagnosis results.

[0057] This embodiment also provides a low-altitude remote sensing monitoring system for maize leaf spot disease in saline-alkali land, the system comprising:

[0058] Data acquisition module: Periodically acquires multi-source remote sensing images of the target cornfield, and preprocesses the multi-source remote sensing images to generate a multi-channel orthophoto map covering the target cornfield;

[0059] Weed pathogen identification module: Based on the multi-channel orthophoto image, weeds in the target cornfield are segmented into instances, and disease diagnosis is performed on the segmented weed instances to identify one or more pathogenic weeds;

[0060] Spatiotemporal latent risk map generation module: Based on the geographical coordinates of the one or more pathogenic weeds, combined with historical meteorological data within a preset time window, a spatiotemporal latent risk map covering the target cornfield is generated;

[0061] Corn leaf spot disease detection module: Based on the spatiotemporal latent risk map, the target cornfield is divided into multiple areas with different risk levels, and a differentiated detection network and adaptive preference decision function that match the risk level are used to detect the corn plants in the area.

[0062] Monitoring module: Generates comprehensive monitoring and risk warning maps or variable spray prescription maps that include corn disease diagnosis results.

[0063] This embodiment details the specific method for low-altitude remote sensing monitoring of maize leaf spot disease in saline-alkali land.

[0064] In a preferred embodiment, the application scenario of this example is limited to a modern, large-scale cooperative corn planting base located in a plain. The base is a flat, contiguous area covering up to 50 hectares (750 mu), with soil exhibiting moderate to mild salinization. The monitoring period is early spring, when the corn is in the seedling stage (e.g., 3-5 leaf stage). This stage is when weeds begin to proliferate, overwintering pathogens begin to become active, and corn leaf spot disease is highly prevalent. Due to the vast planting area, it is difficult for personnel and traditional agricultural machinery to enter the plots, making manual inspection impossible.

[0065] The core task of multi-source remote sensing data acquisition and flight path planning is to periodically acquire multi-source, high-precision image data that accurately reflects the health status of the corn canopy using a UAV remote sensing platform through preset flight and shooting parameters, so as to provide a high-quality data foundation for subsequent disease identification and risk modeling.

[0066] To balance endurance efficiency for large-scale operations with takeoff and landing flexibility, this embodiment uses a vertical takeoff and landing fixed-wing UAV as the flight platform. This platform integrates at least one visible light (RGB) camera module, which is a full-frame aerial survey camera with a sensor effective pixel count of no less than 45 million and equipped with a high-quality fixed-focus lens. The flight platform should also integrate a multispectral camera module, which is an integrated agricultural multispectral camera containing at least five independent sensor channels for acquiring images in the blue light (center wavelength 450nm), green light (center wavelength 560nm), red light (center wavelength 650nm), red-edge (center wavelength 730nm), and near-infrared (840nm) bands. The red-edge and near-infrared bands are extremely sensitive to vegetation health, chlorophyll content, and changes in cell structure, and are key data sources for distinguishing between salt stress and fungal diseases. The flight platform should also integrate a high-precision positioning and attitude system. The platform integrates a real-time dynamic differential (RTK-GPS) module and a high-precision inertial measurement unit (IMU) to ensure that the geographic coordinate accuracy of the image data reaches the centimeter level and the attitude angle accuracy is better than 0.1 degrees.

[0067] Considering that the cycle from infection to the appearance of visible symptoms and the production of the next generation of spores for the pathogen of maize leaf spot under suitable environmental conditions (temperature 20-32℃, relative humidity above 90%) is approximately 72 hours, and that meteorological conditions (especially wind and rainfall) are key variables driving spore dispersal, an excessively long time span would include multiple complex and conflicting meteorological events, rendering risk modeling based on historical weather meaningless. Conversely, overly frequent data collection (e.g., once daily) would lead to unnecessary data redundancy and operational costs.

[0068] Therefore, this embodiment sets a data acquisition cycle. The cycle is 3 days. This cycle can effectively capture disease before it is triggered by a single meteorological event and completes an epidemic cycle, ensuring that the historical meteorological data used for risk analysis (e.g., the past 72 hours) is strongly correlated with the actual development stage of the disease in the field, thus achieving a balance between monitoring timeliness and economy.

[0069] To ensure the integrity and high quality of data collection, a parallel "S"-shaped flight path covering the entire field was planned. Ground sampling distance. The pixel size in an image represents the actual size of the ground feature and is a core indicator for measuring the image's detail. To ensure clear differentiation of early lesions on corn leaves at the millimeter level, a sufficiently small pixel size needs to be set. .

[0070]

[0071] in Ground sampling distance, in centimeters per pixel (cm / pixel). This represents the drone's flight altitude relative to the corn canopy. Considering the potential uneven growth of corn in saline-alkali soil, it is set at 30 meters. (cm) to acquire high-resolution images while ensuring safety. The pixel size of the camera sensor is 4.4 micrometers in this embodiment, which is the pixel size of the RGB camera. cm). The focal length of the camera lens is 35 mm in this embodiment. cm).

[0072] To ensure the success rate of subsequent image stitching and the accuracy of the 3D model, a sufficient heading overlap rate needs to be set. and lateral overlap rate Among them, the heading overlap rate is set. Set the lateral overlap rate. This overlap ratio ensures that each ground point is covered by at least 5-7 images from different angles, guaranteeing the generation of high-quality, distortion-free orthophoto maps.

[0073] Based on the set parameters, the route spacing is calculated. and photo interval .

[0074] route spacing The calculation formula is:

[0075] in, For the sensor width, in this embodiment the RGB camera sensor size is 36mm x 24mm, therefore cm.

[0076] Photo interval The calculation formula is:

[0077]

[0078] in, For the sensor height, in this embodiment cm. To ensure image quality, the fixed-wing cruise speed is set to 15 meters per second for the drone's flight speed. (cm / s).

[0079] Based on the above parameter settings and calculations, the UAV will periodically complete the multi-source remote sensing data collection task for 50 hectares of cornfield according to the planned route. The collected data has high spatial resolution, high positioning accuracy, and rich spectral information, and the collection cycle is closely coupled with the disease epidemic pattern, laying a solid foundation for all subsequent analysis steps.

[0080] Next, in this embodiment, the collected discrete multi-source UAV images will be preprocessed and transformed into a multi-channel orthophoto map with accurate geographic coordinates and radiometric measurements covering the entire operation area.

[0081] A single image captured by a drone only covers a limited area and suffers from perspective distortion (objects appear larger when closer and smaller when farther away) due to the lens angle. For the large-scale farmland of 50 hectares addressed in this application, several discrete images must be stitched together to form a complete orthophoto map. This orthophoto map undergoes orthorectification to eliminate perspective distortion and pixel displacement caused by terrain undulations, ensuring that all features (corn, weeds, soil) within the map have a uniform scale and accurate geographic coordinates. This facilitates accurate area statistics, spatial relationship analysis, and the generation of topographic maps that can be directly used by agricultural machinery.

[0082] Specifically, individual discrete UAV imagery, each with its own independent pixel coordinate system and subject to perspective distortion, cannot be directly and accurately calculated for distance, azimuth, and area. Without stitching, the true spatial relationship between a pathogenic weed in image A and a corn plant in image B (e.g., 3.5 meters apart, located northeast of the corn plant) cannot be calculated. Only by stitching together a single orthorectified georeferenced map can each pixel on the map possess unique, true latitude and longitude coordinates. This allows us to place all spatial elements, such as pathogens, healthy plants, and risk models, on the same coordinate plane for computation, forming the basis for all spatial analysis steps in this invention.

[0083] Meanwhile, to achieve full coverage of the cornfield area to be inspected, a high repetition rate is typically set during image acquisition. This means that any corn stalk in the field may appear in 5 to 10 or even more different images. Analyzing each image independently would not only result in duplicate calculations of the same plant, wasting significant computational resources, but more seriously, subtle variations in shooting angle, lighting, and shadows could lead to inconsistent identification of the plant (e.g., it might be identified as diseased in 3 images but healthy in another 4). This would cause significant confusion in the final decision. The generation process of orthophoto maps, through the principle of multi-view geometry, integrates information from multiple perspectives, generating a unique, standardized, vertically top-down view for each object in the field. This is equivalent to constructing a single source of fact, fundamentally eliminating data redundancy and inconsistencies caused by observation angles, ensuring the stability and reliability of subsequent analysis.

[0084] However, it should be clarified that a complete orthophoto map is not used as a single input for detection, but as a globally unified spatial information database with precise geographic coordinates. The value of a complete orthophoto map lies in its macroscopic global spatial benchmark, rather than as a whole unit of analysis.

[0085] In performing disease detection and identification, this invention employs an overlapping slicing strategy. Specifically, along the XY coordinate axes of the orthophoto image, it is cut into a large number of standard image blocks of uniform size. For example, it is cut into image blocks of 1024×1024 pixels.

[0086] Because environmental factors such as light intensity, solar angle, and atmospheric conditions change during drone flight, the pixel brightness values ​​(DN values) recorded in the raw images cannot directly represent the true reflectivity of ground objects. To ensure the comparability of data collected at different times and in different areas, radiometric correction is necessary.

[0087] This embodiment employs a real-time calibration method based on a downlink illumination sensor (DLS). The DLS is integrated with a multispectral camera to measure the solar irradiance of the current environment in real time during flight. The DN value of each multispectral image is converted into physically meaningful surface reflectance using a radiometric calibration model. :

[0088]

[0089] in, For ground features in the band Surface reflectivity. For the sensor to receive in the band Radiance Real-time measurements by DLS in the band Downward irradiance.

[0090] Through real-time correction by the downlink illumination sensor (DLS), all multispectral images are normalized to a uniform reflectance scale, which reduces the interference of illumination changes on spectral analysis.

[0091] During the image stitching process, the canopies of corn and weeds are not rigid, stationary targets. They can sway and deform due to slight winds, resulting in slight differences in vegetation morphology within overlapping areas of adjacent aerial images. Standard image stitching algorithms (such as SIFT) are prone to generating numerous mismatches in such weakly rigid scenarios, leading to ghosting, blurring, or tearing in the final stitched result. To address this, this invention employs a two-level feature matching strategy.

[0092] The two-level feature matching strategy includes global rigid transformation estimation based on surface stability features:

[0093] Global rigid transformation estimation of stable features utilizes invariant elements in the scene, such as exposed soil surfaces, to calculate initial, robust geometric transformation relationships between adjacent images.

[0094] For each RGB image A pre-trained lightweight semantic segmentation network is applied to segment the region into vegetation region masks. and soil area mask .

[0095] The ORB algorithm is used only for masking soil regions. Feature point detection and descriptor generation are performed within a defined range. A highly reliable set of matching points is obtained by brute-force matching of soil region feature points in adjacent images combined with RANSAC to remove mismatched point pairs. ,in and These are soil feature points with the same name in the two images.

[0096] Based on highly reliable matching point sets Calculate the affine transformation matrix that describes the overall position and orientation relationship between the two images. Affine transformation matrix Used to correct translation, rotation, and scaling caused by changes in the attitude of the drone.

[0097]

[0098] The two-level feature matching strategy also includes vegetation non-rigid deformation compensation based on dense optical flow.

[0099] In addition to global alignment, it is also necessary to address the local, non-rigid displacement of the vegetation canopy caused by wind.

[0100] Original image Through matrix Perform the transformation to obtain the pre-aligned image. To make it consistent with the image Roughly aligned on the coordinate system.

[0101] exist and Overlapping areas, only for vegetation area masks Within each pixel, a displacement vector is calculated for each vegetation pixel using dense optical flow. This displacement vector describes the displacement caused by canopy oscillation from To the pixel-level movement of its corresponding position in the new image.

[0102] Building from raw images any point in To its precise position in the coordinate system of the final stitched image mapping function :

[0103]

[0104] The mapping function First, a global affine transformation is performed. Point Rigid alignment is performed, and then a non-rigid compensation vector calculated from the dense optical flow field is superimposed on top of this. .

[0105] Utilize optimized mapping relationships A 3D point cloud of the cornfield was reconstructed using multi-view stereo matching. Based on this point cloud, a digital surface model describing the undulations of the canopy top surface was generated. Using the DSM as the elevation benchmark, each original image underwent pixel-by-pixel orthorectification to eliminate all geometric distortions caused by topographic relief and perspective projection.

[0106] The final result is a single, multi-channel GeoTIFF orthophoto map. Each pixel in this map has unique geographic coordinates and a set of attribute values ​​(R, G, B, Red-Edge, NIR reflectance), providing a highly consistent and accurate data foundation for all subsequent analysis steps.

[0107] Next, this embodiment details the pixel-level instance segmentation of all weeds in the image. Facing challenges such as dense, occluded, irregularly shaped, and diverse-scale weeds in a saline-alkali soil background, this embodiment employs a progressively optimized segmentation network to achieve progressive optimization from coarse to fine regions.

[0108] The progressive optimization segmentation network includes a feature extraction backbone network, a feature pyramid network, a candidate region generation network, and a dual-branch parallel processing and progressive optimization network.

[0109] The input to the progressive optimization segmentation network is a sliced ​​image block. ,in Enter the height and width dimensions (e.g., 1024). This represents the number of channels.

[0110] Feature extraction backbone network (Preferred, ResNet-50) Extract a set of multi-scale feature maps.

[0111] in, For the first The output feature map of the layer, .

[0112] Feature map Feed into the feature pyramid network Generate a feature pyramid that integrates multi-scale information. .

[0113]

[0114] Candidate Region Generation Network (RPN) Applying this to the feature pyramid generates a set of coordinates for candidate regions. ,in For the first The bounding box coordinates of the candidate regions.

[0115] For each candidate region Execute a parallel processing flow. The parallel processing flow consists of a context bootstrap header, an instance mask header, and an optimization module.

[0116] In cornfields, weeds are among the few targets. If a single-stage algorithm that outputs the global segmentation map all at once is used, the model is easily overwhelmed by the dominant corn features, resulting in insufficient learning and recognition capabilities for these few targets. This invention employs a two-stage segmentation method: first generating candidate regions, then performing fine-grained segmentation. The candidate region generation network performs initial detection, quickly identifying all suspected weed targets in the entire image, significantly narrowing the scope of subsequent processing. All subsequent modules only perform calculations on these high-value candidate regions, achieving precise allocation of computational resources.

[0117] The context header Used to generate a global semantic context graph .

[0118] It is a fully convolutional network whose input is the highest resolution feature map. Contextual Guide Header Defined as:

[0119]

[0120] in, This indicates a kernel size of convolutional layers, For batch normalization layer, This is the ReLU activation function. The structure contains four consecutive... Conv-BN-ReLU block, finally composed of a Convolutional layers reduce the number of channels to (Vegetation / Soil), these three channels correspond to three semantic categories: 1) Corn, 2) Weeds, 3) Background / Soil.

[0121] After passing through the Softmax function, this module outputs a pixel-level semantic probability map for three-class classification. Context-guided head loss function Using pixel-level cross-entropy:

[0122]

[0123] in, for The size of the feature map For pixels True semantic tags, for In pixels This belongs to the category The predicted probability.

[0124] In a green field, if only a local area is considered, the algorithm struggles to distinguish whether the green leaves in front of it belong to corn or weeds. The context guidance head generates a global semantic map with three categories: corn, weeds, and background, providing crucial prior knowledge for the subsequent fusion module. When the network focuses on a weed obscured by corn leaves, the global map guides the fusion module: "The current segmentation target is a weed, and the object obscuring it has a semantic category of corn." This semantic category-based distinction is the most effective way to solve the problem of overlapping objects of the same color across different categories. It elevates the segmentation problem from simple morphological differentiation to semantically guided morphological differentiation, significantly reducing segmentation ambiguity.

[0125] Example of a dock For each candidate region Generate an initial coarse mask .

[0126] Example of a pier For each Operation via RoIAlign From the feature pyramid Extracting feature maps of fixed size .

[0127]

[0128] Example of a pier Processing to form an initial coarse mask :

[0129]

[0130] in, It represents 4 consecutive The network body is composed of Conv-BN-ReLU blocks. It is a step size of 2. Deconvolution layer Reduce the number of channels to 1. It is a sigmoid activation function that outputs a pixel-level probability mask. .

[0131] The instance masking module prioritizes speed and localization over high quality. It provides a rough, initial mask outline for each candidate region. This module is only responsible for identifying the approximate location and shape of the weed and outputting a low-resolution, coarse mask. While maintaining extremely high computational efficiency, the most time-consuming and challenging task of handling boundary details is delegated to the more powerful BFR module. This avoids investing excessive resources in immature, fine-grained calculations at the initial stage, ensuring the overall efficiency of the process.

[0132] Context instance feature fusion module Inject global context information into instance features.

[0133] Extract instance masking dock The output is used as instance features. Simultaneously, from the penultimate feature map of the context guide, the corresponding context features are extracted using RoIAlign. .

[0134]

[0135] in, This indicates that the spliced ​​features are initially fused and dimensionality reduced. Indicates an expansion rate of Convolutional layer. This indicates splicing along the channel dimension. This is the final fused convolutional layer.

[0136] The parallel multi-branch dilated convolutional structure in the context instance feature fusion module enables the fusion process to simultaneously consider spatial relationships at different scales. For example, the branch with a dilation rate of 1 focuses on the fine texture of the weed leaves themselves, while the branch with a dilation rate of 5 allows the current pixel to see the contextual features belonging to corn further away. This multi-scale information fusion ensures that the output feature map contains both instance-specific features and contextual semantic features at each pixel, providing rich and effective criteria for accurately distinguishing between corn and weeds.

[0137] Optimization module In cascade mode ( Iterative optimization of the mask region. Each BFR unit The structure is a residual module that receives the feature map from the previous stage. And output the optimized feature map. and mask update diagram .

[0138]

[0139]

[0140]

[0141] area mask for:

[0142]

[0143] in, For pixel coordinates, For mask The set of outline pixels, Euclidean distance. This is the distance threshold.

[0144] No. The final refined mask of the stage From the previous stage mask Updated graph of current forecast Composed of:

[0145]

[0146] in, For bilinear interpolation upsampling, This is pixel-level multiplication.

[0147] BFR module loss Calculations are performed only in the contour region, allowing the network to focus on the boundary points:

[0148]

[0149] in, It is a set of pixels representing the boundary region defined by the actual mask. It is an indicator function.

[0150] Traditional segmentation methods optimize the entire mask with equal intensity, which is a huge waste of computation. The BFR module defines the most uncertain boundary region in the prediction result of the previous stage, and in the next stage, focuses all the network's computing power on re-predicting this boundary loop at a higher resolution. Cascading allows this optimization to be carried out iteratively, with each subsequent stage continuing to correct based on the previous stage, and the resolution increasing step by step.

[0151] Next, this example performs a preliminary diagnosis of maize leaf spot disease on each segmented weed instance and outputs its location and disease confidence score. Disease diagnosis within segmented weed areas presents unique challenges compared to conventional image classification: the initial symptoms of maize leaf spot on weeds are typically millimeter-sized chlorotic spots or necrotic foci, with subtle features. Simultaneously, salt stress can also cause physiological traits in weed leaves such as yellowing, scorching, and necrotic spots. These traits are extremely similar to the pathological symptoms of leaf spot in visible light images, making reliance solely on RGB texture features highly susceptible to false positives. Even within the segmented weed mask, there is interference from complex backgrounds such as changes in lighting, shadows, and mud / water deposits.

[0152] To overcome the above challenges, this embodiment designs a dual-stream heterogeneous feature fusion diagnostic network, which can simultaneously make a comprehensive judgment on the health status of weeds from two dimensions: morphological texture and physiological and biochemical aspects.

[0153] For each output weed instance mask Extracting corresponding image slices from multi-channel orthophoto maps . Includes RGB and multispectral co- Data from each channel. This diagnostic network. The input is .

[0154] The dual-stream heterogeneous feature fusion diagnostic network processes RGB data and multispectral data respectively.

[0155] Visual Networks Used to extract the morphological, color, and texture features of lesions. Image slices The RGB three channels are for visual networks Input. Sensory network A lightweight convolutional neural network is employed, preferably using the first four stages of EfficientNet-B0 as the feature extractor. After global average pooling, the output is a feature vector containing visual information. ,in This refers to the feature dimension of the output of this stream.

[0156] Multispectral Physiological Feature Stream:

[0157] Multispectral network Extracting spectral characteristics that reflect the intrinsic physiological health of vegetation is key to distinguishing between disease and salt stress. Multispectral network Image slices It has two bands: red-edge and near-infrared.

[0158] The normalized red-edge vegetation index, which is highly sensitive to vegetation stress, is calculated based on the input bands. .

[0159]

[0160] in, and These are the reflectance values ​​after radiometric correction in the near-infrared and red-edge bands, respectively. The NDRE index is sensitive to changes in chlorophyll content. Small leaf spot lesions lead to chlorophyll degradation, resulting in significant local low values ​​on the NDRE plot, while salt and alkali stress usually manifests as a more widespread and uniform decrease in NDRE values.

[0161] The data from three channels—the original red-edge, near-infrared band, and the calculated NDRE index—are fed into a shallow convolutional network. After global average pooling, the network outputs a feature vector containing physiological state information. ,in This refers to the feature dimension of the output of this stream.

[0162] Simply combine the two feature vectors and Simply concatenating features doesn't allow the model to autonomously determine which feature is more important in a specific situation. This embodiment introduces a channel attention fusion module. .

[0163] Concatenate the two feature vectors along the dimension: The concatenated vector The data is fed into a multilayer perceptron (MLP), which consists of two fully connected layers, to learn an attention weight vector along one channel dimension. .

[0164]

[0165] in, and It consists of two fully connected layers. This is the ReLU activation function.

[0166] The original concatenated feature vector is multiplied element-wise with the learned attention weight vector to obtain the final fused feature vector. .

[0167]

[0168] The fused feature vector Send to sorting head The classification head It consists of a fully connected layer and a softmax activation function, and outputs a two-dimensional probability vector. .

[0169]

[0170] in, and This is an example of the weed. Confidence level for being diagnosed as healthy and having small spots.

[0171] Output a list of pathogenic weeds. Each entry in the list of pathogenic weeds contains: one weed diagnosed as diseased (e.g., The precise geographic coordinates of the weed instances (calculated from the centroid of their masks); their disease confidence level. .

[0172] During the non-growing season, besides residing in diseased plant debris in the soil, a crucial survival strategy for the pathogen of maize leaf spot is parasitism on various grassy weeds. These weeds act as refuges for the pathogen, helping it survive the winter or fallow period. When spring arrives and field conditions are suitable, these weeds often germinate earlier than maize, allowing the pathogen to spread and multiply on them first, forming the primary source of infection in the field. Infected weeds act as breeding and transmission bases for the pathogen within the maize field. While the maize is still in its relatively vulnerable seedling or growth stage, they continuously release large amounts of pathogen spores into the surrounding environment. These spores are easily spread to neighboring maize plants through wind and rain, completing the infection from weeds to maize and triggering the initial disease outbreak in the maize field. Therefore, controlling the disease on weeds largely prevents the disease from spreading widely in the field. Traditional methods often wait until visible lesions appear on corn leaves before diagnosis, by which time the disease has already occurred, and control efforts are already reactive. This invention, by monitoring weeds as early warning systems, captures subtle signs of a large-scale outbreak of disease on corn. Even if the corn appears perfectly healthy at that moment, finding lesions on weeds allows for the scientific prediction of the most likely infected areas using risk models.

[0173] Next, this embodiment transforms the discrete, confirmed point-source infection information into a continuous, quantified spatiotemporal latent risk map covering the entire cornfield. This risk map aims to simulate and predict the spread and deposition distribution of pathogenic spores released by pathogenic weeds in the field over a past period, thereby scientifically identifying high-risk areas where the most likely infected corn plants are located.

[0174] The list of pathogenic weeds is defined as follows: .

[0175]

[0176] in, This represents the total number of confirmed pathogenic weeds. Each element in the list... It is a tuple containing its geographic coordinates and disease confidence score:

[0177] in, It is the first The precise geographical coordinates of the centroid of the pathogenic weed. It is the first Confidence level of disease incidence in individual pathogenic weeds.

[0178] Define the epidemiological time window Based on the analysis of the course of corn leaf blight, this embodiment sets... To start from the current monitoring time Backwards 72 hours, and discretized in hours, i.e. .

[0179] Based on the geographical extent of the target cornfield, a request is made to the meteorological data service provider via an application programming interface (API) to obtain data within the specified time window. Within this area, gridded historical meteorological data covers a specific region. For each discrete moment... Meteorological parameters that need to be obtained include: hourly average wind field vector. That is, including the average wind speed of that hour. Wind direction. Hourly average relative humidity. Hourly cumulative rainfall .

[0180] The spread of pathogenic spores does not equate to successful infection. Spores need to remain active to spread through the air, and after reaching the host surface, they require suitable temperature and humidity conditions to germinate and invade. Therefore, this embodiment constructs a time-series infection efficiency weighting function. Used to quantify historical moments Under certain meteorological conditions, the overall efficiency of spores from release to successful infection.

[0181] in: This is a humidity effect function used to describe the influence of humidity on spore activity. When the humidity is below a certain threshold, spores are easily inactivated.

[0182]

[0183] in, It is the humidity threshold for spores to remain active. It is a coefficient that controls the steepness of the curve.

[0184] This is a rainfall effect function used to describe the impact of rainfall on spore deposition and germination. Light or drizzling rain is beneficial for spore attachment and germination, while heavy rain may wash away the spores.

[0185] This embodiment uses a piecewise function for modeling:

[0186]

[0187] in, To determine the optimal rainfall threshold, To control the attenuation coefficient of the scouring effect of heavy rain.

[0188] In terms of pathogens In a local coordinate system with the origin at the windward direction as the x-axis, any point At any moment because Spore concentration produced It can be represented as:

[0189]

[0190] in, It is the first The spore release rate of a pathogen per unit time. This example correlates it with diagnostic confidence. Related: , This is a basic release rate constant. It is a moment The average wind speed. The Pasquale-Gifford diffusion coefficient represents the standard deviation in the crosswind (y) and vertical (z) directions, respectively, and represents the downwind distance. An empirical function of atmospheric stability level. The atmospheric stability level is determined by referring to tables based on the day's solar radiation and wind speed. The effective source height for spore release is set to the average height of the weed canopy in this embodiment. The target receiving height is set to the average height of the corn canopy in this embodiment.

[0191] Any grid cell in the entire cornfield Total latent risk value It is its entire time window Inside, received from all sources of infection. The sum of spore concentrations after infection efficiency weighting.

[0192]

[0193] in: Global coordinates In terms of pathogens Using the origin as the time The wind direction is the relative coordinate in the local coordinate system of the x-axis.

[0194] Internal summation Indicates at time ,point Receive from all The sum of spore concentrations of each pathogen. External summation. This indicates the cumulative concentration of effective infecting spores for each hour over the past 72 hours.

[0195] Total latent risk values ​​are calculated on each grid covering the entire cornfield, generating a spatiotemporal latent risk map. The spatiotemporal latent risk map is a heatmap, showing the value of each pixel. After normalization.

[0196] The spread of diseases is a dynamic process involving both time and space. The spores released from lesions found on weeds today may have already reached the other side of the field yesterday or the day before, carried by wind and rain. Therefore, simply knowing the current location of the pathogenic weeds is far from sufficient; it is necessary to trace the possible transmission routes of the pathogen over the past few days.

[0197] This embodiment is designed entirely based on the fundamental principles of disease epidemiology. Instead of employing simple concentric circles or static sectors centered on the pathogen, it introduces the concept of an epidemiological time window. This time window is not arbitrarily set, but rather corresponds to the life cycle of the corn leaf blight pathogen from infection to the production of the next generation of spores. By retrospectively analyzing historical meteorological data that actually occurred within this window period, this method reconstructs a dynamic and continuous history of pathogen spread, which is closer to reality than static models based on a single moment.

[0198] Pathogenic spores are not lifeless dust particles; their spread, survival, and infection success rates are highly dependent on environmental conditions. A purely physical diffusion model might tell us where the spores drifted to, but it cannot determine whether the spores that drifted there are still alive and capable of causing disease.

[0199] In this embodiment, the temporal infection efficiency weights are designed to assign a biologically effective weight to the physical diffusion process at each historical moment. For example, the model can identify a significant difference between dry, windy nights (spores spread far but are easily inactivated, resulting in low weights) and humid afternoons with light rain (spores may spread short distances but are highly likely to survive and infect, resulting in high weights). In this way, the present invention does not simply simulate the trajectory of spores, but rather their effective infection trajectory, which allows the final risk map to exclude interference from events where effective infection is impossible.

[0200] Furthermore, the infection risk faced by corn in a particular area of ​​the field is not the result of a single event, but rather the cumulative effect of multiple pathogens from different directions and multiple transmission events over different time periods. This embodiment ultimately generates a spatiotemporal latency risk map, which is a cumulative infection pressure map, by summing and accumulating the effective infective spore concentrations generated by all pathogens at all time points. Areas with higher values ​​on the map represent areas that received a greater total amount of effective spores in the past 72 hours, and thus have a higher probability of infection.

[0201] Unlike existing technologies that employ a uniform, indiscriminate scanning method across the entire field, this embodiment utilizes an adaptive preference-based detection and diagnosis method based on risk map partitioning. This method divides the field into different levels of regions according to the risk map, then matches different performance detection networks to different regions, and finally uses an adaptive preference-based decision function to intelligently judge the detection results. This achieves an optimal balance between efficiency and accuracy on a macroscopic level, while simultaneously improving the reliability of the diagnosis.

[0202] Performing a comprehensive, high-precision disease scan of a 50-hectare cornfield is computationally extremely expensive and unnecessary. Disease occurrence and spread exhibit a high degree of spatial clustering. Therefore, concentrating limited, high-quality computational resources on high-risk areas predicted by risk models is an efficient strategy.

[0203] This embodiment uses a dual-threshold segmentation method for region division.

[0204] Set two global risk thresholds: a high-risk threshold. and low risk threshold For risk diagrams Each grid cell in Based on its risk value It is classified into one of the following three non-overlapping geographical regions:

[0205] High-risk areas All satisfied of This area is adjacent to and primarily downwind of pathogenic weeds, making it extremely prone to disease outbreaks.

[0206] medium-risk area All satisfied of The collection of [elements]. This area is a potential diffusion zone.

[0207] low-risk areas All satisfied of This area comprises the majority of the farmland and has a low probability of disease occurrence.

[0208] Output three geographic zoning masks covering the entire field, which will guide the subsequent differential detection process.

[0209] Different risk areas have different requirements for detection tasks. High-risk areas require zero tolerance for missed detections, while low-risk areas require rapid scanning to avoid false positives. To this end, this embodiment pre-trains and deploys two detection network models with different performance and complexity.

[0210] Refined detection network Application area is and A deep learning model with powerful feature extraction capabilities is employed, preferably Vision Transformer (ViT-Base) or EfficientNet-B5. A refined detection network is then implemented. Used to process high-resolution image slices and identify early lesions. Refined detection network. Its advantages are high accuracy and low false negative rate, while its disadvantages are large computational load and slow inference speed.

[0211] Coarse detection network Application Area A lightweight and efficient model is employed, with MobileNetV3-Large being the preferred choice. The detection network is coarsened. It prioritizes reasoning speed. Its advantage is its extremely high speed, enabling rapid scanning of large areas. Its disadvantage is that it is not sensitive to fine lesions.

[0212] The raw output of a neural network is a probability value (e.g., the probability that this pixel is a lesion is 0.55). The final classification as diseased or healthy depends on a pre-defined decision threshold. A fixed threshold cannot adapt to complex field conditions. In high-risk areas, even if the model only gives a low probability of 0.55, combined with the high-risk prior of the surrounding environment, it should be more likely to be considered a lesion (aggressive strategy); conversely, in low-risk areas, a probability of 0.55 is likely noise or an artifact, and it should be more likely to be considered not a lesion (conservative strategy).

[0213] The embodiment designs an adaptive decision threshold function. The output value of this function is dynamic; it is directly related to the currently detected value. Latent risk value Related.

[0214]

[0215] in, It is aimed at The decision threshold is calculated dynamically. This is a baseline threshold, representing the judgment standard when there is no prior information (i.e., a moderate risk value of 0.5), preferably... =0.6, It is a preference factor used to control the strength of preference. In this embodiment, it can be set . yes Normalized risk value.

[0216] For any one in the field According to its region ( Select the corresponding detection network ( or By reasoning, the original predicted probability of the presence of a lesion at that point is obtained. .

[0217] Applying adaptive preference decision functions Make the final judgment:

[0218]

[0219] in, It is a point The final diagnostic category.

[0220] In high-risk areas (e.g.) The decision threshold will decrease:

[0221] .

[0222] This means that even with refined networks Even if only a vague signal slightly above 0.52 is given, the system will still identify it as a disease and adopt an aggressive, erroneous strategy to ensure the highest disease detection rate.

[0223] In low-risk areas (e.g.) The decision threshold will increase:

[0224] .

[0225] This means that the coarse network Only when a very strong confidence signal above 0.68 is given will the system identify it as a disease and adopt a prudent strategy that minimizes false alarms.

[0226] Effective monitoring reports need to present complex, multi-dimensional information to farm managers in the most intuitive and easily understandable way.

[0227] The system will generate a multi-layered GIS-compatible file (preferably in GeoTIFF format), which uses the generated multi-channel orthophoto as a base map and overlays the following vector and raster layers in sequence:

[0228] Spatiotemporal latent risk layer: the generated risk map As a semi-transparent thermal layer overlaid on the base map, it visually displays the distribution of potential risks in the field.

[0229] Pathogen Weed Location Layer: Outputs a list of pathogenic weeds. Each pathogen coordinates Mark specific vector points (preferably red dots) and indicate their disease confidence level. .

[0230] Diagnostic corn lesion layer: Final diagnostic category For all grid cells with disease The data is then aggregated. To facilitate observation, adjacent diseased units are clustered to form one or more disease patches. Each disease patch is stored and visualized as a vector polygon object.

[0231] Traditional plant protection operations involve spraying pesticides indiscriminately and uniformly across the entire field, which not only wastes pesticides and increases costs but can also cause environmental pollution and crop damage. The ultimate goal of this invention is to achieve precise variable-rate pesticide application. Therefore, monitoring results are converted into operational instructions, i.e., prescription maps, that can be directly read and executed by plant protection drones or intelligent ground spraying equipment.

[0232] Based on risk zoning and the final diagnostic results, each grid unit of the field will be... Calculate an accurate pesticide application decision instruction. This instruction contains two key parts: pesticide type and application dosage.

[0233] For high-risk areas The area within which lesions have been confirmed to exist, and a buffer zone extending outwards from it. Within the designated area, the decision-making instruction is to use a therapeutic disinfectant.

[0234] For high-risk areas Areas with undiagnosed lesions, the entire medium-risk area And a secondary buffer zone extending beyond the therapeutic area, with the decision directive to use a lower-cost protective bactericide.

[0235] For low-risk areas The decision-making instruction was not to administer the medicine.

[0236] The dosage is not fixed but dynamically calculated based on the overall severity index of the area. For any grid cell requiring application... Its CSI value It is determined by both the latent risk value and the density of lesions.

[0237]

[0238] in, It is a latent risk value. This is the pixel density of lesions in the local neighborhood (i.e., the proportion of pixels diagnosed as diseased out of the total number of pixels). If there are no lesions in the neighborhood, this value is 0. It is a weighting factor used to balance the importance of latent risk and existing disease status.

[0239] Recommended dosage (Unit: ml / m²) Calculated using a linear mapping function:

[0240]

[0241] in, and These are the minimum and maximum effective application doses set in agronomic principles.

[0242] For each grid cell Calculated pesticide type and dosage The file is encoded in a standard prescription map file format (preferably SHP or ISO-XML). This file contains geographic coordinate information and corresponding operational instructions, which can be directly loaded and executed by automated equipment such as agricultural drones.

[0243] The system will also automatically generate a quantitative analysis report in PDF format, including: monitoring date, field number, maize growth stage, total field area, areas of high, medium, and low risk zones and their percentage of the total area, statistics on the confirmed disease area and number of lesions in each risk zone, estimated pesticide usage, and the savings compared to indiscriminate spraying across the entire field.

[0244] This embodiment also provides a low-altitude remote sensing monitoring system for maize leaf spot disease in saline-alkali land, the system comprising:

[0245] Data acquisition module: Periodically acquires multi-source remote sensing images of the target cornfield, and preprocesses the multi-source remote sensing images to generate a multi-channel orthophoto map covering the target cornfield;

[0246] Weed pathogen identification module: Based on the multi-channel orthophoto image, weeds in the target cornfield are segmented into instances, and disease diagnosis is performed on the segmented weed instances to identify one or more pathogenic weeds;

[0247] Spatiotemporal latent risk map generation module: Based on the geographical coordinates of the one or more pathogenic weeds, combined with historical meteorological data within a preset time window, a spatiotemporal latent risk map covering the target cornfield is generated;

[0248] Corn leaf spot disease detection module: Based on the spatiotemporal latent risk map, the target cornfield is divided into multiple areas with different risk levels, and a differentiated detection network and adaptive preference decision function that match the risk level are used to detect the corn plants in the area.

[0249] Monitoring module: Generates comprehensive monitoring and risk warning maps or variable spray prescription maps that include corn disease diagnosis results.

[0250] An electronic device includes a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the aforementioned low-altitude remote sensing monitoring method for maize leaf spot disease in saline-alkali land.

[0251] A computer-readable storage medium storing a computer program that, when executed by a processor, implements the aforementioned low-altitude remote sensing monitoring method for maize leaf spot disease in saline-alkali land.

[0252] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium, and when executed, it can include the processes of the embodiments of the above methods. Any references to memory, storage, databases, or other media used in the embodiments provided in this application can include non-volatile and / or volatile memory. Non-volatile memory can include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memory can include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in various forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), dual data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link DRAM (SLDRAM), Rambus direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM), etc.

[0253] In this specification, the same or similar parts between the various embodiments can be referred to mutually. Each embodiment focuses on describing the differences from other embodiments. In particular, the descriptions of the embodiments described later are relatively simple, and relevant parts can be referred to the descriptions of the foregoing embodiments.

[0254] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A low-altitude remote sensing monitoring method for maize leaf spot disease in saline-alkali land, characterized in that, The method includes: Periodically acquire multi-source remote sensing images of the target cornfield, and preprocess the multi-source remote sensing images to generate a multi-channel orthophoto map covering the target cornfield; Based on the multi-channel orthophoto image, weeds in the target cornfield are segmented into instances, and disease diagnosis is performed on the segmented weed instances to identify one or more pathogenic weeds. Based on the geographical coordinates of the one or more pathogenic weeds, and combined with historical meteorological data within a preset time window, a spatiotemporal latent risk map covering the target cornfield is generated. Based on the spatiotemporal latent risk map, the target cornfield is divided into multiple areas with different risk levels, and a differentiated detection network and adaptive preference decision function that match the risk level are used to detect diseases in the corn plants within the area. Generate a comprehensive monitoring and risk warning map or a variable spraying prescription map that includes the diagnostic results of corn diseases; Disease diagnosis was performed on the segmented weed instances using a dual-stream heterogeneous feature fusion diagnostic network, which processed RGB data and multispectral data respectively. The dual-stream heterogeneous feature fusion diagnostic network includes a visual network and a multispectral network, and processes the feature vectors output by the visual network and the multispectral network through a channel attention fusion module. and The feature vectors are then fused to obtain the final fused feature vector. The calculation method is as follows: in, and These are the feature vectors output by the visual network and the multispectral network, respectively. This indicates splicing along the channel dimension; This is a channel-dimensional attention weight vector; Construct a time-series infection efficiency weight function Used to quantify historical moments Under specific meteorological conditions, the overall efficiency of spore release from release to successful infection is defined as: in, and They are time points Hourly average relative humidity and hourly cumulative rainfall; It is a humidity effect function used to describe the effect of humidity on spore activity; It is a rainfall effect function used to describe the influence of rainfall on spore deposition and germination; Generated Spatiotemporal Latent Risk Map any one of its grid cells Total latent risk value In order to be within the preset epidemiological time window Inside, received from all The weight of the pathogenic weeds and their time-series infection efficiency. The weighted sum of spore concentrations: in, For at any time By the Disease-causing weeds The concentration of spores produced; It is a grid cell In terms of pathogens Using the origin as the time The wind direction is a relative coordinate in the local coordinate system of the x-axis; The adaptive preference decision function used includes an adaptive decision threshold function. The output value of this function is related to the current detected point. Latent risk value Related, which is defined as: Final diagnostic category : in, As a baseline threshold, As a preference factor; To detect the network's response to the currently detected point The original predicted probability of the presence of lesions.

2. The method for low-altitude remote sensing monitoring of maize leaf spot disease in saline-alkali land according to claim 1, characterized in that: The multi-source remote sensing images are preprocessed using a two-level feature matching strategy, which constructs a feature matching matrix from the original images. any point in To its precise position in the coordinate system of the final stitched image mapping function Its definition is: in, This is the global affine transformation matrix. It was calculated by matching only the soil surface stability feature points between adjacent images; For non-rigid compensation vectors, the Through the The dense optical flow is calculated for the pixels in the vegetation area within the overlapping region of the initially aligned images.

3. The method for low-altitude remote sensing monitoring of maize leaf spot disease in saline-alkali land according to claim 2, characterized in that: Weeds are segmented into instances using a progressively optimized segmentation network, which includes an optimization module that iteratively optimizes the mask boundaries in a cascaded manner. In the Each iteration stage leads to the final refined mask. From the previous stage mask And the currently predicted mask update graph It is composed of combinations, and the combination method is as follows: in, and The first Stage and First Refined masking for each stage; It is in the The phase specifically targets the high-resolution mask update map predicted for the boundary region; This is a bilinear interpolation upsampling operation. Pixel-level multiplication; To use the mask from the previous stage The calculated boundary region mask is defined as the mask The Euclidean distance between the center and the contour pixel set is less than or equal to a preset distance threshold. A set of pixels.

4. The method for low-altitude remote sensing monitoring of maize leaf spot disease in saline-alkali land according to claim 3, characterized in that: The progressive optimization segmentation network includes a context guide head and an instance masking head; the context guide head is used to generate a global semantic context graph that distinguishes corn, weeds and background soil; the instance masking head is used to generate an initial coarse mask for each candidate region; Before iterative optimization, a context instance feature fusion module is also included, which is used to fuse the instance features of the instance masking tool. Contextual features of the context header Perform fusion to generate a fused feature map. .

5. A method for low-altitude remote sensing monitoring of maize leaf spot disease in saline-alkali land according to claim 1, characterized in that: The generated variable spray prescription map contains grid cells for each pesticide application. Calculated recommended dosage : in, and These are the set minimum and maximum effective dosages, respectively. This is the overall severity index for that grid cell, which is determined by the latent risk value of that grid cell. and the pixel density of the lesion in its local neighborhood We obtain the weighted sum.

6. A low-altitude remote sensing monitoring system for maize leaf spot disease in saline-alkali land, implementing the low-altitude remote sensing monitoring method for maize leaf spot disease in saline-alkali land as described in any one of claims 1-5, characterized in that, The system includes: Data acquisition module: Periodically acquires multi-source remote sensing images of the target cornfield, and preprocesses the multi-source remote sensing images to generate a multi-channel orthophoto map covering the target cornfield; Weed pathogen identification module: Based on the multi-channel orthophoto image, weeds in the target cornfield are segmented into instances, and disease diagnosis is performed on the segmented weed instances to identify one or more pathogenic weeds; Spatiotemporal latent risk map generation module: Based on the geographical coordinates of the one or more pathogenic weeds, combined with historical meteorological data within a preset time window, a spatiotemporal latent risk map covering the target cornfield is generated; Corn leaf spot disease detection module: Based on the spatiotemporal latent risk map, the target cornfield is divided into multiple areas with different risk levels, and a differentiated detection network and adaptive preference decision function that match the risk level are used to detect the corn plants in the area. Monitoring module: Generates comprehensive monitoring and risk warning maps or variable spray prescription maps that include corn disease diagnosis results.

Images