Field crop seedling condition diagnosis method and device based on unmanned aerial vehicle remote sensing image

Abstract

The invention discloses a field crop seedling condition diagnosis method and device based on an unmanned aerial vehicle remote sensing image, belongs to the field of agricultural information technology and crop phenotype detection, and particularly relates to field crop seedling condition diagnosis based on the unmanned aerial vehicle remote sensing image. The problems of low seedling condition diagnosis efficiency, poor accuracy and difficulty in supporting accurate seeding quality evaluation caused by the deficiencies of data acquisition normalization, farmland structure adaptability of a model and analysis dimension comprehensiveness in the prior art are solved. The method comprises a data acquisition step, a plot ridge identification step and a ridge seedling identification step. The field crop seedling condition diagnosis method and device based on the unmanned aerial vehicle remote sensing image are suitable for intelligent agricultural scenes such as automatic growth monitoring and sowing quality evaluation of field crops in the seedling stage and precise field management (such as variable fertilization and precise reseeding).

Description

Technical Field

[0001] This invention relates to the fields of agricultural information technology and crop phenotypic detection, and in particular to the diagnosis of field crop seedling conditions based on UAV remote sensing images. Background Technology

[0002] Accurate and efficient understanding of crop emergence status in the field (such as emergence rate and uniformity of seedling spatial distribution) is an important foundation for implementing precision field management, evaluating the quality of sowing operations, and predicting crop yield, and is of key significance to the high-quality development of modern agriculture.

[0003] Traditional monitoring of crop seedling conditions in field crops mainly relies on manual field surveys conducted by agricultural technicians. This method is not only inefficient and highly subjective, but also has limited coverage, making it difficult to meet the needs of large-scale, high-frequency, and quantitative seedling condition diagnosis. In addition, the survey process can easily cause mechanical damage to seedlings.

[0004] In recent years, with the development of remote sensing technology, macro-level agricultural monitoring methods based on satellite platforms have been applied. However, satellite remote sensing images have low spatial resolution, and seedling monitoring based on them is usually limited to the population level, making it difficult to achieve high-precision identification and positioning at the scale of individual seedlings, and thus failing to meet the needs of precision agriculture for fine management at the individual level.

[0005] Unmanned aerial vehicle (UAV) remote sensing technology offers a new solution for near-field remote sensing monitoring in agriculture due to its flexibility, efficiency, and relatively low cost. By equipping itself with multiple sensors, UAVs can quickly acquire high-resolution image data of farmland. Based on this, combined with image processing techniques or deep learning-based target detection, crop seedlings can be identified and located from the images, significantly improving the automation level of seedling condition information extraction. For example: Chinese patent document CN202110705485.4 discloses a method for monitoring seedling growth using depth images obtained from UAV remote sensing and based on image grayscale value analysis. While this type of method achieves a certain degree of automated analysis, its core relies on traditional image processing techniques. It has poor adaptability to complex field environments (such as changes in light intensity, soil background interference, and seedling shading), and its feature extraction capabilities are limited. Consequently, the accuracy and robustness of the monitoring results often fail to meet the requirements of precision agriculture.

[0006] Chinese patent document CN202510210769.4 discloses a method for detecting seedlings in the field by improving the YOLOv8 model. This type of method utilizes convolutional neural networks to automatically learn seedling characteristics, resulting in a significant improvement in detection accuracy compared to traditional methods. Nevertheless, existing seedling monitoring technologies based on UAV remote sensing still have significant shortcomings, limiting their widespread application in large-scale production practices. First, at the data acquisition level, there is a lack of standardized and efficient UAV data acquisition specifications for crop condition diagnosis tasks. Existing methods mostly rely on autonomous UAV navigation for simple regional image acquisition, or emphasize only high forward and lateral overlap rates for later image stitching, failing to fully consider the targeted adaptation of flight altitude, flight path, and the actual geometry of the field ridges. This easily leads to problems such as regional overlap and inconsistent scale in the acquired data, and changes in flight altitude can cause image distortion, not only reducing acquisition efficiency but also introducing additional complexity and errors to subsequent image processing (such as stitching, geometric correction, and partitioning).

[0007] Secondly, at the level of intelligent information extraction, many existing seedling detection models fail to fully utilize the unique spatial structure knowledge of farmland (such as row orientation and row spacing). They often directly input the entire image into the model for general target detection, which is easily affected by complex factors such as weeds between rows, soil background, and shadows. In scenarios with dense seedlings or occlusion, the false positive and false negative rates are relatively high.

[0008] Furthermore, in terms of analytical dimensions and decision support, existing technologies are mostly limited to simple statistics on seedling quantity or inversion of single growth indicators, failing to form a complete sowing quality assessment system. For key seedling conditions directly determined by sowing operation quality, such as the uniformity, evenness, and degree of seedling deviation, as well as the presence of gaps in the rows, there is a lack of systematic definitions, extraction, and comprehensive evaluation methods. This makes it difficult to directly translate technological achievements into effective decision-making basis for guiding field operations such as replanting and thinning.

[0009] Therefore, there is an urgent need for a complete technical solution that can achieve efficient and standardized data collection, precise seedling information extraction combined with field structure knowledge, and comprehensive evaluation of multi-dimensional sowing quality, in order to overcome the shortcomings of existing technologies and comprehensively improve the efficiency, accuracy and practicality of field crop seedling diagnosis. Summary of the Invention

[0010] This invention proposes a method and device for diagnosing crop seedling conditions in field crops based on UAV remote sensing images. It addresses the shortcomings of existing technologies in areas such as data acquisition standardization, model adaptability to farmland structure, and comprehensiveness of analysis dimensions. These shortcomings lead to low efficiency, poor accuracy, and difficulty in supporting precise sowing quality assessment in seedling diagnosis. The method for diagnosing crop seedling conditions in field crops based on UAV remote sensing images, as described in this invention, includes the following steps: Data acquisition steps: acquire multiple plot images of fields awaiting seedling condition diagnosis based on UAV remote sensing technology; preliminarily identify abnormal areas in each plot image; for the preliminarily identified abnormal areas, acquire point maps of the abnormal areas based on UAV remote sensing technology, and obtain preliminary seedling shortage diagnosis data for the abnormal areas; wherein, the resolution of the point maps of the abnormal areas is higher than that of the plot images. Cellular ridge identification steps: A pre-trained DeepLabV3+ model is used as the ridge segmentation model; each cell image is input into the ridge segmentation model to obtain a binary mask of the ridge body, where each pixel is labeled as either the ridge body or the background; a geometric processing algorithm is used to filter out the background from the binary mask of the ridge body, retaining the ridge body pixel region, and calculating the ridge body boundary coordinate sequence; based on the ridge body boundary coordinate sequence, a closed polygonal region is generated as the ridge body region, and an image of the identified ridge body region is obtained; The seedling identification steps on the ridge are as follows: The pre-trained YOLOv11n model is used as the seedling detection model on the ridge; the image of the identified ridge area is input into the seedling detection model on the ridge, and the individual seedlings are detected to obtain the position of each seedling in the ridge area and the actual number of seedlings, and to construct an effective seedling sequence.

[0011] Furthermore, in a preferred embodiment, the cell image is acquired using the following method during the data acquisition step: Select fields at the emergence stage based on the target crop type as fields for seedling condition diagnosis; Drones were used to patrol the fields for crop condition diagnosis in order to collect images of the fields and the crops on them, and to obtain multiple images of the plots. During field patrol operations: The flight path of the drone is planned parallel to the direction of the field where the crop condition needs to be diagnosed: the drone flies at a constant speed along the direction of the ridge to collect images, ensuring that the images collected by each drone are not repeated and that all images collected by the drones cover the entire area of ​​the field where the crop condition needs to be diagnosed; the actual field area corresponding to each collected image is equal, and the actual field area corresponding to a single collected image is defined as a small area. The drone collects data at the same altitude relative to the ground.

[0012] Furthermore, a preferred embodiment is provided, wherein in the data acquisition step, abnormal regions in each cell image are initially identified; for the initially identified abnormal regions, a point map of the abnormal regions is acquired based on UAV remote sensing technology, and preliminary seedling shortage diagnosis data for the abnormal regions is obtained; wherein the resolution of the point map of the abnormal regions is higher than that of the cell images, including the following: For each cell image, vegetation index and canopy coverage are detected. Areas with vegetation index and canopy coverage below a given vegetation threshold are marked and pinpointed according to their coordinates to delineate abnormal areas. The anomaly region mapping was collected using the following method: The drone flew directly above all the marked points and began acquiring images of the abnormal area; The drone flies at a constant speed along the direction of the ridge to collect images of abnormal areas; the direction of the ridge extension when collecting images of abnormal areas is the same as the direction of the ridge extension when collecting images of the plot. The drone's acquisition altitude relative to the ground is the same, and the acquisition altitude when acquiring images of abnormal areas is lower than the acquisition altitude when acquiring images of the cell. The seedling emergence status is inverted by using the collected point maps of abnormal areas to obtain preliminary seedling loss diagnosis data for abnormal areas.

[0013] Furthermore, a preferred embodiment is provided, wherein the ridge identification step employs a geometric processing algorithm to filter out the background from the ridge binary mask, retain the ridge pixel region, and calculate the ridge boundary coordinate sequence, including the following steps: Steps for extracting the centerline of a ridge: Obtain the probability map of the binary mask of the ridge; assuming that the ridges are arranged in rows and columns in the probability map, and the image width direction is taken as the column direction; where the predicted probability of the ridge in the k-th row and j-th column is... , ; Calculation along the column direction satisfies The weighted average of the pixel column coordinates is used as the column coordinates of the centerline of the ridge to obtain the centerline of the ridge: ; in, represents the column coordinates of the center line of the k-th row of the ridge; W is the image width of the probability plot, in pixels; This is an indicator function; I = 1 when the condition inside the parentheses is true, and I = 0 otherwise. Segmentation threshold, ; Row spacing calculation steps: Based on the centerline column coordinates of two adjacent rows of rows, calculate the spacing between the two adjacent rows of rows, and take the average spacing as the pixel-level average row spacing. ; Where d is the average row spacing; N is the number of rows; Scale conversion steps: Convert pixel-level average row spacing to physical-scale row spacing: ; in, Pixel-to-physical scale conversion coefficient; The steps for extracting the ridge direction baseline are as follows: In the pixel coordinate system, the coordinates of the center line column of the obtained ridge are detected. The Hough transform in polar coordinate form is used to transform the detection of the ridge direction baseline into peak detection in the parameter space. The angle distribution of the candidate ridge direction baselines obtained by detection is statistically analyzed. After removing outliers, the average of the high-frequency angles is taken as the ridge direction baseline angle. Based on the ridge direction baseline angle, three types of ridge direction are divided: horizontal ridge direction, vertical ridge direction, and oblique ridge direction. The steps for calculating ridge width are as follows: A series of uniformly distributed sampling points are set at certain intervals along the centerline of the ridge; at each sampling point, the binary mask of the ridge is scanned to the left and right sides along a direction perpendicular to the ridge reference angle; when a pixel value changes from the ridge to the background, it is determined as the left or right boundary point of the ridge; all sampling points are traversed to obtain a series of ordered boundary point pairs consisting of left and right boundary points, which serve as the ridge boundary coordinate sequence; the pixel distance between the left and right boundary points of each sampling point is calculated, and the average value is taken as the local ridge width; the average ridge width of all ridges is calculated by averaging the local ridge widths to obtain the average ridge width of the farmland.

[0014] Furthermore, in a preferred embodiment, the method further includes: The plot analysis steps are as follows: Based on the location of each seedling in the obtained ridge area, the actual number of seedlings, and the effective seedling sequence, the sowing quality in the ridge area is evaluated by judging uniformity, deviation, evenness, and missing seedlings and broken rows. in: Uniformity: The uniformity of seedling distribution on the ridge is reflected by the uniformity of spacing and the uniformity of position; the uniformity of spacing is used to reflect the consistency of the distance between adjacent seedlings; the uniformity of position is used to reflect the degree of concentration of seedlings on the center line of the ridge. Degree of deviation: indicates the degree of lateral deviation of the growth position of a single seedling relative to the center line of the ridge; Uniformity: Indicates the consistency of the spacing between adjacent seedlings on the ridge, reflecting the uniformity of planting density; Missing seedlings: This indicates that one or more seedlings are missing; Broken rows: This indicates that multiple seedlings are missing from a row, forming a long blank section.

[0015] Furthermore, in a preferred embodiment, the method further includes: Field analysis steps: Based on the location of each seedling in the obtained ridge area, the actual number of seedlings, and the effective seedling sequence, calculate the seedling planting density of each plot and evaluate the straightness of the ridge.

[0016] This invention also proposes a field crop seedling condition diagnosis device based on UAV remote sensing images, the device comprising the following modules: Data acquisition module: acquires multiple plot images of fields awaiting seedling condition diagnosis based on UAV remote sensing technology; performs preliminary identification of abnormal areas in each plot image; for the preliminarily identified abnormal areas, acquires a point map of the abnormal area based on UAV remote sensing technology, and obtains preliminary seedling shortage diagnosis data for the abnormal area; wherein, the resolution of the point map of the abnormal area is higher than that of the plot image. Cellular ridge identification module: A pre-trained DeepLabV3+ model is used as the ridge segmentation model; each cell image is input into the ridge segmentation model to obtain a binary mask of the ridge, where each pixel is labeled as either the ridge or the background; a geometric processing algorithm is used to filter out the background from the binary mask of the ridge, retaining the ridge pixel region, and calculating the ridge boundary coordinate sequence; based on the ridge boundary coordinate sequence, a closed polygon region is generated as the ridge region, and an image of the identified ridge region is obtained; The seedling identification module on the ridge: The pre-trained YOLOv11n model is used as the seedling detection model on the ridge; the image of the identified ridge area is input into the seedling detection model on the ridge to detect individual seedlings, obtain the position of each seedling in the ridge area and the actual number of seedlings, and construct an effective seedling sequence.

[0017] The present invention also proposes a computer device comprising: a processor and a memory, the memory being used to store executable instructions of the processor, the processor being configured to execute the above-described field crop seedling condition diagnosis method based on UAV remote sensing images by executing the executable instructions.

[0018] The present invention also proposes a computer storage medium storing a computer program, wherein when the computer program is executed, it performs the field crop seedling condition diagnosis method based on UAV remote sensing images described above.

[0019] The present invention also proposes a computer program product, including a computer program / instruction, which, when executed by a processor, implements the steps of the field crop seedling condition diagnosis method based on UAV remote sensing images described above.

[0020] The present invention has the following beneficial effects: 1. The field crop seedling condition diagnosis method based on UAV remote sensing images described in this invention achieves high-quality image data acquisition in complex farmland environments by designing a new standard that integrates ridge orientation recognition, contour flight, and fixed-point collection. This method avoids data redundancy and geometric distortion caused by different flight altitudes and overlapping paths in traditional methods, thus laying a reliable data foundation for subsequent accurate analysis.

[0021] 2. The field crop seedling diagnosis method based on UAV remote sensing images described in this invention achieves progressively refined analysis from coarse localization of "row area" to precise detection of "single seedling" by constructing a multi-stage analysis process that coordinates the DeepLabV3+ row segmentation model and the YOLOv11n seedling detection model. This effectively overcomes the problem of missed and false detection of seedling targets in complex backgrounds and significantly improves the automation level and recognition accuracy of seedling information extraction.

[0022] 3. The field crop seedling condition diagnosis method based on UAV remote sensing images described in this invention introduces the key agronomic spatial structure of "ridges" as the core analysis dimension and defines quantitative agronomic indicators such as uniformity, deviation, and evenness. This enables a multi-dimensional and refined assessment of sowing quality, surpassing the limitations of traditional methods that only count the number of seedlings. It provides a direct and reliable scientific basis for precise farmland management and agricultural decision-making (such as replanting and fertilization).

[0023] 4. The field crop seedling condition diagnosis method based on UAV remote sensing images described in this invention achieves a balance between rapid scanning of large areas of farmland and detailed diagnosis of problem areas by adopting a hierarchical detection strategy of "general survey (normal altitude flight) + detailed survey (low-altitude fixed point of abnormal areas)". While ensuring the overall inspection efficiency, it can accurately locate abnormal areas such as missing seedlings and weak seedlings and obtain high-resolution detailed images, which greatly improves the pertinence and depth of diagnosis.

[0024] 5. The field crop seedling condition diagnosis method based on UAV remote sensing images described in this invention significantly enhances the model's perception and adaptation to agricultural scene-specific features (such as crop row structure and small seedling targets) by performing targeted structural optimization and training strategy adjustments on general models such as DeepLabV3+ and YOLOv11n (e.g., adjusting ASPP void ratio, adapting seedling anchor boxes, and adopting CIoU loss function). This improves the robustness and practicality of the entire technical solution in real field environments.

[0025] The method and device for diagnosing the seedling condition of field crops based on UAV remote sensing images described in this invention are applicable to smart agriculture scenarios such as automated growth monitoring, sowing quality assessment, and precision field management (such as variable fertilization and precision replanting) of field crops (such as rice, wheat, corn, soybeans, etc.) during the seedling stage. Attached Figure Description

[0026] To more clearly illustrate the technical solutions in the embodiments of the present invention, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0027] Figure 1 This is a flowchart illustrating a method for diagnosing crop seedling conditions in field based on UAV remote sensing images, as one embodiment of the present invention. Figure 2 In one embodiment of the present invention, a schematic diagram of the ridge binary mask obtained by inputting a cell image into a ridge segmentation model is shown; the ridge binary mask is displayed in the form of a ridge semantic segmentation image, wherein the black and gray part represents the background and the dark red lines represent the ridge (or ridge row). Figure 3 This is a schematic diagram of the effective seedling sequence obtained by the seedling detection model on the ridge, as one embodiment of the present invention. Detailed Implementation

[0028] To make the technical solutions and advantages of the present invention clearer, the specific embodiments of the present invention will be described in further detail and completely below with reference to the accompanying drawings. The various embodiments described below are only some preferred embodiments of the present invention, and not all of them; the various embodiments described below are intended to explain the present invention and should not be construed as limiting the present invention; reasonable combinations of the technical features defined in the various embodiments of the present invention, as well as all other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort, are all within the scope of protection of the present invention.

[0029] Implementation Method 1: A method for diagnosing crop seedling conditions in field crops based on UAV remote sensing images, the method comprising the following steps: Data acquisition steps: acquire multiple plot images of fields awaiting seedling condition diagnosis based on UAV remote sensing technology; preliminarily identify abnormal areas in each plot image; for the preliminarily identified abnormal areas, acquire point maps of the abnormal areas based on UAV remote sensing technology, and obtain preliminary seedling shortage diagnosis data for the abnormal areas; wherein, the resolution of the point maps of the abnormal areas is higher than that of the plot images. Cellular ridge identification steps: A pre-trained DeepLabV3+ model is used as the ridge segmentation model; each cell image is input into the ridge segmentation model to obtain a binary mask of the ridge body, where each pixel is labeled as either the ridge body or the background; a geometric processing algorithm is used to filter out the background from the binary mask of the ridge body, retaining the ridge body pixel region, and calculating the ridge body boundary coordinate sequence; based on the ridge body boundary coordinate sequence, a closed polygonal region is generated as the ridge body region, and an image of the identified ridge body region is obtained; The seedling identification steps on the ridge are as follows: The pre-trained YOLOv11n model is used as the seedling detection model on the ridge; the image of the identified ridge area is input into the seedling detection model on the ridge, and the individual seedlings are detected to obtain the position of each seedling in the ridge area and the actual number of seedlings, and to construct an effective seedling sequence.

[0030] In this embodiment, spatial coordinate correlation and fusion are performed between the cell image and the (corresponding) anomaly area fixed-point map: By using the pixel coordinate system of the image, the "global region label" (identified ridge area, effective seedling sequence) of the low-resolution image (cell image) is bound to the "local detail data" (preliminary seedling loss diagnosis data of the abnormal area) of the high-resolution image (abnormal area pinpoint map) to form a complete diagnostic result of "region + detail".

[0031] In this embodiment, data acquisition of cell images and anomaly area pinpoint maps is as follows: By using a layered data acquisition approach of "low-resolution global coarse screening + high-resolution local focusing," we have achieved both efficient global scanning of the field and accurate diagnosis of abnormal areas. This balance between efficiency and accuracy is an innovative approach that distinguishes us from traditional single-resolution solutions.

[0032] In this embodiment, UAV remote sensing technology is an airborne remote sensing technology that integrates unmanned aerial vehicles, remote sensing sensors, telemetry and remote control technology, communication technology, GPS differential positioning system, and data processing technology. Using the UAV as an aerial platform, remote sensing sensors (such as high-resolution CCD cameras, multispectral imagers, and small digital cameras) acquire information, and a computer processes the image information to create images according to certain accuracy requirements.

[0033] In this embodiment, the cell images collected by UAV remote sensing technology include multi-source geospatial data such as the UAV's own position, the position of the target being photographed, altitude, and attitude angle (camera attitude information).

[0034] The geospatial data can be summarized as georeferenced data accompanying the cell images, for example: Drone GPS / RTK positioning data: provides the spatial location (latitude, longitude, and altitude) at the moment of camera exposure.

[0035] Camera intrinsic parameters: describe the optical characteristics of the lens (such as focal length and distortion) and are used to correct geometric distortions in images.

[0036] Ground control points (GCPs): Known coordinate markers placed within the field to serve as absolute references for optimizing positioning accuracy.

[0037] In this embodiment, the DeepLabV3+ model is a semantic segmentation network with an encoder-decoder structure. Its encoder is based on a deep convolutional neural network and introduces a spatial pyramid pooling module (ASPP) to extract multi-scale features in parallel using atrous convolutions at different ratios, effectively capturing the contextual information of the ridges (capturing contextual information in a way that uses multiple sampling rates and multiple receptive fields to obtain high-level semantic features). Its decoder gradually recovers and refines the boundary information of the ridges by fusing the high-level semantic features output by the encoder with the detailed features of the shallow layers of the network.

[0038] In this implementation, the DeepLabV3+ model is improved as follows: To address the characteristics of ridge features in agricultural drone images, the dilation rate (or simply dilation rate) of the dilated convolution in the ASPP module was specifically adjusted, and the model input layer was adapted to the actual resolution of field images (such as plot images). This improvement aims to enhance the model's ability to extract ridge morphological features, especially in distinguishing ridges from soil and seedlings in complex field backgrounds, while ensuring the best balance between model computational efficiency and recognition performance.

[0039] Furthermore, its decoder effectively improves the accuracy of segmentation boundaries through a structured upsampling process. Specifically, the decoder first performs bilinear interpolation upsampling (e.g., 4 times) on the high-level semantic features output by the encoder to achieve the same spatial resolution as the low-level feature maps output by the intermediate layers of the network backbone. Then, the two types of features are fused and concatenated, and convolutional layers are used for feature refinement. Finally, the refined features are subjected to another bilinear interpolation upsampling (e.g., 4 times) to restore the original input image (cell image) size, ultimately outputting pixel-level prediction results. This decoding design significantly optimizes the final segmentation mask, especially the accuracy of the target object (ridge) edges, by reusing detailed information from shallow features.

[0040] Bilinear interpolation is an upsampling algorithm, an image magnification method based on local pixel weighted averaging. It is computationally efficient and can achieve a relatively smooth transition effect.

[0041] The original upsampling method of the DeepLabV3+ model is to directly upsample the encoding result by a factor of 16.

[0042] By introducing an encoder structure with bilinear interpolation upsampling, instead of a one-time upsampling, a two-step process is adopted: first, the feature map output by the encoder is upsampled by 4 times, then fused with low-level features from the shallow layers of the network, and then upsampled by 4 times again to finally restore the original image size. By fusing with low-level features containing rich position and texture information, the object boundary details lost due to network downsampling are effectively repaired, resulting in a more accurate segmentation contour. This improvement significantly enhances the model's segmentation accuracy at object boundaries, making the predicted map more consistent with the real label, and achieving an upgrade from "coarse segmentation" to "fine segmentation".

[0043] The DeepLabV3+ model's improvement over existing technologies lies not in inventing bilinear interpolation itself, but in designing a novel encoder-decoder architecture. This architecture creatively utilizes the decoder path and selectively fuses it with low-level features from the encoder at specific stages (such as after upsampling), thus systematically solving the boundary ambiguity problem in semantic segmentation. The key is the fusion with low-level features; the decoder not only performs upsampling but, more importantly, concatenates the features with those from early layers of the network backbone (such as ResNet). While these low-level features are not semantically strong, they retain more precise edge and positional information, serving as a crucial source for achieving refined boundary details.

[0044] In this embodiment, the DeepLabV3+ model is trained using the following method: Construct a ridge semantic segmentation dataset (training set). The dataset is a pixel-level labeled dataset based on field images of crops in the seedling stage, with the label categories being "ridge", "background", and "seedling". Using this ridge semantic segmentation dataset, the DeepLabV3+ model was trained end-to-end. During training, key parameters such as the hole rate of the ASPP module were adjusted based on the ridge semantic segmentation dataset and the ridge recognition task.

[0045] By constructing a semantic segmentation dataset for ridges and specifically adapting model parameters such as the hole rate, the pre-trained model obtained at the end has stronger specialization and higher segmentation accuracy for the target crop (such as soybean) in the field environment at a specific growth stage (seedling stage), effectively solving the problem of insufficient adaptability of general models in specific agricultural scenarios.

[0046] In this embodiment, after each cell image is input into the row segmentation model (a pre-trained DeepLabV3+ model), the pre-trained DeepLabV3+ model extracts features from the input cell images during the forward propagation process, as follows: The ASPP module's multi-scale feature fusion formula is used to output C from dilated convolutions with different dilation rates (e.g., 1, 6, 12, 18). r (x) is concatenated with image-level global features G(x) to capture contextual information at different scales; The multi-scale feature fusion formula of the ASPP module is as follows: F=Concat[C1(x), C6(x), C 12 (x), C 18 [(x), G(x)]; The mathematical definition of atrous convolution is:

[0047] Where r is the hole rate (expansion rate), used to expand the receptive field of the convolution kernel while maintaining the number of parameters; w is the weight of the convolution kernel.

[0048] In this embodiment, the YOLOv11n model is a lightweight model.

[0049] In this embodiment, a pre-trained YOLOv11n model is used to output the location and number of seedlings: The image of the identified ridge area is input into the pre-trained YOLOv11n model, which limits the detection range to the ridge area (ROI), eliminates background interference such as soil and weeds, and improves detection efficiency and accuracy. A predefined set of reference boxes (anchor boxes) at different scales are used to match the seedling size distribution (e.g., obtained through cluster analysis of training data); The final detection result output by the model after fine-tuning the anchor box is used as the bounding box coordinates, in the format [x_center, y_center, width, height], which represents the center point coordinates and width and height (normalized pixel values) of the seedling bounding box. The seedling location is uniquely determined by the bounding box coordinates; The confidence level (between 0 and 1) is the confidence score of the model's prediction of the presence of seedlings within the bounding box. A threshold (e.g., 0.5) is typically set for the confidence score to filter out low-confidence detection results. The model also outputs category information: if only a single type of seedling is detected, it outputs a single-category label; if it is necessary to distinguish between crops and weeds, or to distinguish between different types of seedlings, it outputs a multi-category label. After the model infers the ridge region in a single image, it uses non-maximum suppression to remove duplicate detection boxes for the same seedling (retaining the box with the highest confidence). Finally, it counts the number of all independent bounding boxes with a confidence level higher than the threshold, which is the actual number of seedlings.

[0050] In this embodiment, the YOLOv11n model structure and its improvements are as follows: The YOLOv11n model is based on its efficient single-stage detection architecture (Backbone-Neck-Head). Its core improvements aim to accurately adapt it to the detection tasks of seedlings on ridges at multiple crop and growth stages in complex farmland environments. Specific optimizations are as follows: Input layer adapted to farmland image resolution: To enable the model to process farmland images collected by drones, the input layer was specifically adjusted; the input size of the model was adjusted to match the resolution of farmland images (such as community images) collected by drones (e.g., 640x640 pixels) to ensure that image information is not distorted or lost due to scaling during the preprocessing stage, thus preserving more seedling details for subsequent feature extraction.

[0051] Lightweight attention mechanism for neck fusion: A lightweight attention mechanism, such as SimAM (Simple, Parameter-free Attention Module) or other efficient attention modules, is introduced into the feature fusion module of the neck (such as Feature Pyramid Network (FPN) and Path Aggregation Network (PAN). This mechanism can make the model pay more attention to key spatial regions and feature channels related to seedlings during feature fusion without significantly increasing computational complexity, effectively suppressing the interference of complex backgrounds such as soil and shadows, thereby enhancing the feature extraction ability of seedlings, especially partially occluded seedlings.

[0052] The output layer uses the CIoU Loss function: To address the potential for inaccurate localization caused by the small size and dense distribution of seedlings on ridges, CIoU Loss (Complete Intersectionover Union Loss) is used as the loss function in the bounding box regression task of the output layer. Compared to the traditional IoU Loss, CIoU Loss considers three geometric factors simultaneously: overlap area, center point distance, and aspect ratio, which can more effectively guide the model to perform accurate localization and significantly improve the detection accuracy of small-scale seedlings.

[0053] In this embodiment, the YOLOv11n model training method and its improvements are as follows: The model's training strategy is designed closely around the specificities of agricultural application scenarios, aiming to improve the model's generalization ability and practical deployment performance.

[0054] A multi-scenario ridge seedling dataset was constructed: images of ridge seedlings of different crop types (such as corn, soybeans, and wheat), under different light conditions (sunny and cloudy days), and at different growth stages were collected. These images were accurately labeled to generate information including the bounding boxes of seedling locations; the dataset was divided into training, validation, and test sets in a ratio (e.g., 8:1:1) to ensure the objectivity of the evaluation.

[0055] Implement targeted data augmentation: To improve the robustness of the model in variable field environments, data augmentation techniques such as random rotation, scaling, and brightness and contrast adjustment were applied to the training data. These operations simulated various situations that may occur in the field, effectively increasing the diversity of the data, helping the model reduce overfitting, and learn more generalized seedling characteristics.

[0056] Hyperparameter tuning based on the validation set: During model training, instead of using a fixed set of hyperparameters, key hyperparameters (such as learning rate and batch size) are dynamically adjusted and optimized using the validation set. For example, strategies such as cosine annealing may be used to adjust the learning rate to ensure that the model converges stably to the optimal state. This is a process of continuously monitoring the performance of the validation set (such as loss function value, mAP, etc.) and making feedback adjustments, which is a key step in ensuring the final performance of the model.

[0057] Generate anchor boxes adapted to seedlings: Before training begins, cluster analysis is performed on the width and height of all labeled boxes in the training set to generate a set of prior anchor boxes adapted to the typical size features of seedlings; this enables the model to better match the shape of seedling targets in the early stages of training, accelerates convergence and improves detection recall, especially for small-scale seedlings, the detection effect is significantly improved.

[0058] Implementation Method 2: In the data acquisition step, the cell image is acquired using the following method: Select fields at the emergence stage based on the target crop type as fields for seedling condition diagnosis; Drones were used to patrol the fields for crop condition diagnosis in order to collect images of the fields and the crops on them, and to obtain multiple images of the plots. During field patrol operations: The flight path of the drone is planned parallel to the direction of the field where the crop condition needs to be diagnosed: the drone flies at a constant speed along the direction of the ridge to collect images, ensuring that the images collected by each drone are not repeated and that all images collected by the drones cover the entire area of ​​the field where the crop condition needs to be diagnosed; the actual field area corresponding to each collected image is equal, and the actual field area corresponding to a single collected image is defined as a small area. The drone collects data at the same altitude relative to the ground.

[0059] In this embodiment, a suitable emergence period is selected according to the target crop type (such as soybean), and a drone is used to patrol the field where the seedling condition needs to be diagnosed. The drone is equipped with a visible light camera and embedded equipment to collect RGB images of the field crops (i.e., collected images).

[0060] In this embodiment, during field inspection, the drone's flight path is planned parallel to the extension direction of the ridges. Embedded devices are used to detect the ridge direction and fly at a constant speed along the ridge direction to collect field images in a fixed manner, ensuring that the collected images are not repeated and can cover the entire field area. The actual field area corresponding to each collected image is equal, and the actual field corresponding to a single collected image is defined as a small area.

[0061] In this embodiment, during field inspection, the UAV flies at a constant altitude to adapt to the height difference of the plots, ensuring that the acquisition height of all images remains the same (e.g., 15m), and ensuring that the acquired data is not cluttered due to changes in the height of the plots.

[0062] In this embodiment, during field patrol operations (when collecting images of the area), the drone's acquisition height relative to the ground is 15 meters.

[0063] It should be noted that existing technologies for collecting field data (images) using drones are inefficient and lack regulations regarding the drone's flight altitude and mode. Data collection is typically based solely on autonomous drone navigation. This method suffers from problems such as overlapping collection areas, the impact of altitude variations, inaccurate positioning, and long processing times.

[0064] It should be noted that existing technologies for collecting field data (images) using drones also include a forward overlap acquisition method (such as a forward overlap rate of 80% and a side overlap rate of 70%), which requires subsequent image stitching and cropping.

[0065] The data acquisition method described in this embodiment provides specific operating parameters for UAV data acquisition. Furthermore, the images acquired by each UAV are unique, and all UAV images cover the entire field area to be diagnosed, thus reducing the image acquisition duplication rate. This eliminates the need for subsequent processing steps such as image stitching and cropping using image stitching software, greatly saving data acquisition and image processing time.

[0066] Implementation Method 3: In the data acquisition step, abnormal areas in each cell image are initially identified; for the initially identified abnormal areas, a point map of the abnormal areas is acquired based on UAV remote sensing technology, and preliminary seedling shortage diagnosis data for the abnormal areas is obtained; wherein, the resolution of the point map of the abnormal areas is higher than that of the cell images, including the following: For each cell image, vegetation index and canopy coverage are detected. Areas with vegetation index and canopy coverage below a given vegetation threshold are marked and pinpointed according to their coordinates to delineate abnormal areas. The anomaly region mapping was collected using the following method: The drone flew directly above all the marked points and began acquiring images of the abnormal area; The drone flies at a constant speed along the direction of the ridge to collect images of abnormal areas; the direction of the ridge extension when collecting images of abnormal areas is the same as the direction of the ridge extension when collecting images of the plot. The drone's acquisition altitude relative to the ground is the same, and the acquisition altitude when acquiring images of abnormal areas is lower than the acquisition altitude when acquiring images of the cell. The seedling emergence status is inverted by using the collected point maps of abnormal areas to obtain preliminary seedling loss diagnosis data for abnormal areas.

[0067] In this embodiment, vegetation detection equipment (such as DJI Terra) is used to detect the NDVI (vegetation index) and canopy coverage of each cell image; for areas with canopy coverage of less than 30%, the coordinates are used to mark and define abnormal areas.

[0068] In this embodiment, by analyzing NDVI (vegetation index) and canopy coverage, problem areas (abnormal areas) can be accurately located.

[0069] In this embodiment, for abnormal areas, high-resolution low-flying images (abnormal area fixed-point maps) can be acquired for detailed analysis.

[0070] In this embodiment, based on the marked points, the abnormal area is repeatedly surveyed using a drone to collect data, thereby acquiring a higher resolution map of the abnormal area's location. In this embodiment, only the abnormal areas marked with fixed points are captured with high resolution abnormal area fixed point maps. The UAV flies directly above all the marked fixed points to ensure that the ridge direction is consistent with the overall (small area) flight detection ridge direction, which facilitates subsequent comprehensive analysis.

[0071] In this embodiment, when collecting fixed-point maps of abnormal areas, the drone's acquisition altitude is lower than the altitude at which the cell images are collected.

[0072] For example, if the drone's acquisition altitude is 15 meters when collecting images of a residential area, then the drone's acquisition altitude is 5 meters when collecting fixed-point maps of abnormal areas.

[0073] In this embodiment, when collecting fixed-point maps of abnormal areas, the drone collects images of all ridges and areas with missing seedlings to ensure that the location of the ridges and the details of seedling emergence can be clearly identified.

[0074] In this embodiment, based on the collected abnormal area fixed-point map, the seedling situation is reflected to form preliminary seedling shortage diagnosis data for the abnormal area.

[0075] In this embodiment, the resolution of the anomaly area fixed-point map is much higher than that of the cell image. The seedling status can be inferred by using existing conventional technologies (such as DJI Terra and related seedling detection technologies in patent documents such as CN202110705485.4 and CN202510210769.4).

[0076] Alternatively, in one embodiment, the cell image can also be obtained using another method (heading overlap acquisition method): (1) Based on UAV remote sensing technology, the field to be diagnosed with crop growth is subjected to directional overlap acquisition, and multiple single-area field images are collected; The field images of the single area were acquired using the following methods: Select fields at the seedling stage according to the target crop type as fields to be diagnosed on the seedling condition, and place ground markers on the fields; Drones were used to patrol the field to collect images of the field and its crops, resulting in multiple images of individual areas. During field inspections, the drone's flight path is planned parallel to the direction of the ridges on the field to ensure complete imaging of the ridges.

[0077] In this embodiment, ground markers are placed on the field to facilitate subsequent positioning and segmentation (the complete field image is segmented into multiple cell images).

[0078] In this embodiment, during routine field inspection operations, the drone is equipped with a visible light camera to collect RGB images of the field and its crops, i.e., field images.

[0079] In this embodiment, during routine field survey operations, the drone's flight altitude is controlled at 15m, its speed at 4m / s, its heading overlap rate at 80%, and its lateral overlap rate at 70%.

[0080] Flight altitude (15m): Flight altitude directly determines the ground resolution of the image, that is, how large one pixel represents. An altitude of 15 meters can obtain high-resolution images with centimeter-level resolution, making details of seedlings, soil, and weeds clearly distinguishable. At the same time, this altitude also ensures sufficient safety for the drone in agricultural environments.

[0081] Flight speed (4m / s): This speed is a relatively conservative setting, mainly aimed at balancing image sharpness (avoiding image blurring due to excessive movement) and operational efficiency. The lower speed ensures that the drone's displacement is small during camera exposure, resulting in clearer single images and laying the foundation for high-quality stitching later.

[0082] Forward overlap (80%) and lateral overlap (70%): These are key parameters to ensure the successful generation of complete and seamless orthophoto maps in later stages.

[0083] Forward overlap rate refers to the degree of overlap between two adjacent photos along the same flight path. A high overlap rate (80%) ensures that there are enough common feature points between adjacent photos, allowing the software to stitch them together very reliably, especially during the seedling stage when crop features may be relatively uniform.

[0084] Lateral overlap rate refers to the degree of overlap between photos taken along two adjacent parallel flight paths. A high lateral overlap rate (70%) ensures that photos taken along different flight paths can be stitched together well, avoiding gaps or missing data.

[0085] These two parameters work together to provide a sufficient data foundation for subsequent "feature point extraction and matching", which is a prerequisite for generating high-quality orthophotos and 3D models.

[0086] (2) Stitch together multiple single-area field images into a complete field image; perform post-processing on the complete field image to enhance the contrast between seedlings, soil, and weeds; map the pixel coordinates of the complete field image to latitude and longitude coordinates under the standard geographic coordinate system through sensor data fusion and geometric mapping; divide the complete field image into multiple small area images. The process of segmenting a complete field image into multiple cell images includes the following steps: Steps to determine the segmentation strategy: Based on the geometry of the physical plots on the field to be diagnosed and the target precision, calculate the optimal segmentation granularity of the plot. The steps for generating a segmented grid are as follows: Based on the corner points of the fields in the latitude and longitude coordinates of the standard geographic coordinate system, a segmented grid is constructed; the segmented grid is used to divide the complete field image into multiple field plots, ensuring that each field plot is aligned with the physical field; the field corner point refers to the vertex of the boundary of the physical field on the complete field image; Steps for outputting cell images: Project the segmented grid onto the complete field image, crop it to generate independent image files as cell images, and retain the latitude and longitude coordinates in the standard geographic coordinate system.

[0087] In this embodiment, the geometric shape of the field plots to be diagnosed is, for example, a rectangle or an irregular polygon.

[0088] In this embodiment, the target precision is such as the minimum detection scale for seedlings.

[0089] In this embodiment, the optimal segmentation granularity is such that each field plot covers an actual area of ​​5m × 5m.

[0090] In this embodiment, the segmented mesh is a regular quadrilateral mesh or an adaptive vector boundary: Regular quadrilateral grid: Divided evenly like a chessboard, suitable for fields with regular shapes, and the algorithm is simple and efficient.

[0091] Adaptive vector boundary: Generates a mesh that fits the shape based on the actual boundaries of irregular physical fields (such as polygon corners), making full use of the effective area and avoiding the inclusion of non-farmland areas.

[0092] In this embodiment, the segmented grid is projected onto the complete field image (orthophoto) using georegistration parameters.

[0093] This method effectively reduces the data volume of a single image, improving the algorithm efficiency and robustness of field ridge recognition and crop seedling detection. In this embodiment, the latitude and longitude coordinates of each cell image in the standard geographic coordinate system are preserved in the form of geographic labels.

[0094] In this embodiment, by segmenting the complete field image into multiple small images, the amount of data in a single image is effectively reduced, thereby improving the algorithm efficiency and robustness of field ridge recognition and crop seedling detection.

[0095] In this embodiment, although the single-area field images collected by the UAV contain rich information in the forward overlap acquisition method, they have their own limitations and need to be stitched together: Single-area field images have limited coverage and suffer from high central resolution and significant edge distortion. By setting a high overlap rate (e.g., 60%-80%) during shooting and then using specialized software (such as ContextCapture, GlobalMapper, etc.) to stitch them together, hundreds or even thousands of photos can be seamlessly merged into a unified and complete orthophoto image, thereby obtaining an overall view of a large area.

[0096] The stitching process is more than just simple image pasting. Based on the location and attitude information of the photos, it uses algorithms to perform geometric correction, eliminating image distortion caused by changes in drone flight attitude and lens distortion, generating an image map with a uniform scale to improve data consistency and usability. The final result is an image with precise geographic coordinates, which can be directly overlaid onto existing map base maps for measurement, analysis, and planning.

[0097] In addition, weather and lighting conditions may change during field inspections, leading to inconsistencies in tone and brightness between different batches of photographs. This issue can be resolved through color balancing during the stitching process, resulting in a more harmonious final image that is easier to interpret.

[0098] In this embodiment, the forward overlap acquisition method stitches together multiple single-area field images into a complete field image: Mainstream software used for stitching together image points includes Metashape, Pix4D, and DJI Terra. These tools, based on computer vision and photogrammetry principles, synthesize overlapping single-area field images into a complete field image (orthophoto). The technical process is as follows: Feature point extraction and matching: First, feature points (such as corner points and texture edges) of each single-area field image are detected by algorithms (such as SIFT, SURF or ORB), and then feature descriptors are used for matching to establish the correspondence between single-area field images; for example, the deep learning method combining SuperPoint and SuperGlue can improve the matching accuracy, especially suitable for multispectral images; Alignment and geometric correction: Camera position and pose are estimated based on matching points (optimized by Bundle Adjustment), while lens distortion (based on camera intrinsics) and projection distortion are corrected to generate sparse point clouds; then, 3D point clouds and digital surface models (DSM) are generated through dense matching. Orthophoto generation: Multiple corrected single-area field images are projected onto a unified plane (based on DSM), and fade-in / fade-out or multi-band fusion algorithms are used to eliminate stitching seams, finally outputting a seamless geometrically corrected image, which is the complete field image.

[0099] This process relies on a high overlap ratio set during flight (80% forward overlap and 70% lateral overlap) to ensure reliable matching.

[0100] In this embodiment, the image post-processing in the heading overlap acquisition method includes cropping, rotation, and contrast enhancement. Post-processing aims to improve image quality and enhance the distinguishability of target features (seedlings, soil, weeds), providing a foundation for subsequent analysis.

[0101] Cropping: Define the effective area of ​​farmland based on geographic coordinates or visual boundaries, using the ROI (Region of Interest) cropping function of tools such as OpenCV; used to remove invalid edges in the stitched image (such as the distorted parts in the turning area of ​​a drone) or to focus on key farmland areas, reducing redundant data; Rotation: Rotating the image around its center by a specified angle using affine transformations (such as rotation matrices), often combined with bilinear interpolation to maintain smoothness; used to adjust the image orientation, aligning the direction of the ridges with the coordinate axes (the axes of the pixel coordinate system), facilitating the identification of crop rows and machinery paths. This processing is performed on the stitched digital image (a pixel matrix); the image pixel coordinate system has its origin at the top left corner, with the u-axis (or x-axis) horizontally to the right and the v-axis (or y-axis) vertically downwards; the rotation operation uses algorithms such as affine transformations to rotate the entire pixel matrix, making the direction of the ridges in the image as parallel as possible to the u-axis or v-axis; this greatly simplifies subsequent algorithms, for example, when identifying crop rows, only vertical detection is needed, significantly improving the efficiency and accuracy of algorithms for identifying large ridges and crop seedlings; Mosaic enhancement processing: This technique improves local contrast by enhancing the image, making the spectral or texture differences between seedlings and soil or weeds more significant. It should be noted that seedlings in their early growth stages have similar spectral characteristics to soil and weeds (especially in RGB images), making direct identification susceptible to interference. Enhancing contrast can amplify the differences between categories and improve the accuracy of automated detection.

[0102] Methods for enhancing mosaic effects: Histogram equalization: expands the dynamic range of image grayscale and enhances details in dark areas (such as seedling shadows).

[0103] Color space conversion: Convert RGB to HSV or LAB space, and adjust the saturation / brightness channels separately to highlight green vegetation.

[0104] Gamma correction: Non-linear adjustment of the brightness curve to enhance the sense of depth in low-contrast areas.

[0105] In this embodiment, the heading overlap acquisition method involves geographic coordinate transformation, sensor data fusion, and geometric mapping. This step, which enables precise mapping from pixel coordinates to geographic coordinates (latitude and longitude coordinates in the standard geographic coordinate system), is the core of spatial benchmark construction.

[0106] Sensor data fusion: By integrating multi-source geospatial data (including UAV GPS / RTK positioning data, camera intrinsic parameters (focal length, principal point, distortion coefficient), and ground control point (GCP) measurements), and by unifying the clock and coordinate system, spatiotemporal inconsistencies between sensor data are eliminated. For example, RTK provides centimeter-level UAV position, camera intrinsic parameters define the geometric relationship between pixels and physical space, and GCPs serve as an absolute coordinate reference to correct errors.

[0107] Geometric mapping: This is the process of converting pixel coordinates into geographic coordinates using a camera imaging model (such as a collinearity equation model) and a coordinate transformation chain. For example: Forward intersection: Using camera extrinsic parameters (attitude calculated from GPS / RTK and IMU data) and intrinsic parameters, pixels are projected backward into the object space.

[0108] Coordinate transformation: The object point is gradually transformed from the camera coordinate system to the UAV carrier coordinate system, the global coordinate system (such as WGS84), and finally projected onto the planar coordinate system (such as UTM).

[0109] The core of this geometric mapping is to establish a collinearity equation model in photogrammetry, more specifically: First, using the camera's intrinsic parameter matrix (obtained through calibration), the pixel coordinates (u, v) are converted into normalized direction vectors in the camera coordinate system.

[0110] Subsequently, the extrinsic parameter matrix (i.e., the position and attitude of the camera relative to the UAV carrier) is calculated by combining the GPS / RTK positioning data and IMU attitude data of the UAV, and the direction vector is rotated and translated to the world coordinate system.

[0111] Finally, the three-dimensional coordinates in the world coordinate system are converted into latitude and longitude coordinates and altitude in the target geographic coordinate system (such as WGS84 or CGCS2000) through map projection algorithms (such as Universal Transverse Mercator Projection UTM), thereby achieving a precise one-to-one correspondence between each image pixel and the real geographic location.

[0112] Obtain georeferencing parameters: These are a set of parameters that define the mathematical transformation relationship between pixel coordinates (u, v) and latitude and longitude coordinates in a standard geographic coordinate system. These parameters are typically contained in the metadata of a world file or image.

[0113] In summary, with "sensor data fusion + geometric mapping" as the core, by integrating UAV GPS / RTK positioning data, camera intrinsic parameters, and ground control point (GCP) information, image pixel coordinates are accurately converted into latitude and longitude coordinates under standard geographic coordinate systems such as WGS84 or CGCS2000, realizing a one-to-one correspondence between "pixels and geographic locations" and providing a precise spatial reference for crop growth detection and ROI area geographic positioning.

[0114] In this embodiment, standard geographic coordinate systems, such as WGS84 and CGCS2000 coordinate systems, are used.

[0115] Both WGS84 and CGCS2000 coordinate systems are geocentric coordinate systems, which are global or regional standard geographic coordinate systems.

[0116] In this embodiment, in the heading overlap acquisition method, image segmentation reduces the processing complexity of a single image. Segmenting a complete field image into multiple small plots is a key strategy for optimizing computational efficiency.

[0117] Reason for splitting: Reduce the complexity of processing a single image: Complete field images have high resolution (e.g., hundreds of millions of pixels), and direct processing requires a large amount of memory and computing power. After segmentation, the amount of image data in each cell is reduced, allowing for parallel processing and accelerating the analysis process.

[0118] Improve detection accuracy: Adjust algorithm parameters (such as seedling size threshold) for local areas (such as a single field ridge) to avoid misjudgments caused by global uniform processing.

[0119] Methods for segmenting images into cell blocks: Grid-based rule segmentation: Divides the field image into grids according to a fixed size (e.g., 512×512 pixels), suitable for uniform fields.

[0120] Ground marker-based auxiliary segmentation: Ground markers deployed in the field can be used not only for coordinate correction, but their spatial distribution can also serve as a reference for segmentation boundaries. For example, a vector grid can be generated using ground markers as control points to ensure that each plot corresponds to the actual field partition (such as each experimental field).

[0121] The relationship between field size and detection accuracy: The larger the field, the finer the segmentation needs to be: In order to ensure pixel-level accuracy in seedling detection (such as identifying seedlings at the millimeter level), large fields need to be divided into smaller units to avoid loss of detail due to image scaling.

[0122] The required precision determines the granularity of the segmentation: if it is necessary to identify individual seedlings, the size of the plot should be smaller than the width of the field ridge; if only the growth is being monitored, the size of the plot can be appropriately increased.

[0123] In summary, segmenting the images into smaller cells improves the accuracy of subsequent field ridge recognition and crop seedling detection.

[0124] Implementation Method 4: In the ridge identification step, a geometric processing algorithm is used to filter out the background from the ridge binary mask, retain the ridge pixel region, and calculate the ridge boundary coordinate sequence, including the following steps: Steps for extracting the centerline of a ridge: Obtain the probability map of the binary mask of the ridge; assuming that the ridges are arranged in rows and columns in the probability map, and the image width direction is taken as the column direction; where the predicted probability of the ridge in the k-th row and j-th column is... , ; Calculation along the column direction satisfies The weighted average of the pixel column coordinates is used as the column coordinates of the centerline of the ridge to obtain the centerline of the ridge: ; in, represents the column coordinates of the center line of the k-th row of the ridge; W is the image width of the probability plot, in pixels; This is an indicator function; I = 1 when the condition inside the parentheses is true, and I = 0 otherwise. Segmentation threshold, ; Row spacing calculation steps: Based on the centerline column coordinates of two adjacent rows of rows, calculate the spacing between the two adjacent rows of rows, and take the average spacing as the pixel-level average row spacing. ; Where d is the average row spacing; N is the number of rows; Scale conversion steps: Convert pixel-level average row spacing to physical-scale row spacing: ; in, Pixel-to-physical scale conversion coefficient; The steps for extracting the ridge direction baseline are as follows: In the pixel coordinate system, the coordinates of the center line column of the obtained ridge are detected. The Hough transform in polar coordinate form is used to transform the detection of the ridge direction baseline into peak detection in the parameter space. The angle distribution of the candidate ridge direction baselines obtained by detection is statistically analyzed. After removing outliers, the average of the high-frequency angles is taken as the ridge direction baseline angle. Based on the ridge direction baseline angle, three types of ridge direction are divided: horizontal ridge direction, vertical ridge direction, and oblique ridge direction. The steps for calculating ridge width are as follows: A series of uniformly distributed sampling points are set at certain intervals along the centerline of the ridge; at each sampling point, the binary mask of the ridge is scanned to the left and right sides along a direction perpendicular to the ridge reference angle; when a pixel value changes from the ridge to the background, it is determined as the left or right boundary point of the ridge; all sampling points are traversed to obtain a series of ordered boundary point pairs consisting of left and right boundary points, which serve as the ridge boundary coordinate sequence; the pixel distance between the left and right boundary points of each sampling point is calculated, and the average value is taken as the local ridge width; the average ridge width of all ridges is calculated by averaging the local ridge widths to obtain the average ridge width of the farmland.

[0125] In this embodiment, the geometric processing algorithm aims to transform the pixel-level binary mask of the ridges output by the DeepLabV3+ model into an ordered sequence of vector boundary point coordinates that defines each ridge region.

[0126] In this embodiment, the step of extracting the center line of the ridge is intended to determine the "spine" or geometric axis of each ridge.

[0127] In this embodiment, It can be adjusted according to the field environment, and is generally taken as 0.5.

[0128] In this embodiment, the physical dimension is the spacing between rows. The unit is meters (m).

[0129] In this embodiment, the step of extracting the ridge-direction baseline is as follows: In the image coordinate system, the Hough transform is performed in polar coordinate form. The detection of ridge lines (ridge-oriented baselines) in image space is transformed into the detection of ridge lines in parameter space. Peak detection; statistical detection of the angle values ​​of the ridge lines () The distribution of the angles was analyzed, and after removing outliers, the mean of the high-frequency angles was taken as the reference angle for the ridge direction. ).

[0130] in: In the pixel coordinate system, with the top left corner of the image as the origin (0,0), the x-axis points horizontally to the right, and the y-axis points vertically downward; Hough transform ( It is defined in this coordinate system, for example: Origin: The pixel at the top left corner of the image, with coordinates (0, 0); x-axis: The horizontal direction of the image, with positive to the right; Line: The baseline line along the ridge that needs to be inspected; (Rho): The perpendicular distance from the origin of the image to the line being measured; (Theta): The angle between the perpendicular line drawn from the origin to the measured line and the positive x-axis, with a range of values ​​of ( ).

[0131] In this embodiment, based on Determine the type of monopoly: when At that time, it was a horizontal ridge; when At that time, it is perpendicular to the ridge direction; That When it falls within the other angle range, it is oblique ridge direction.

[0132] In this embodiment, the ridge width calculation steps are as follows: (1) Sampling: A series of evenly distributed sampling points are set at certain intervals along the center line of the ridge; (2) Search: At each sampling point, along the reference angle with the ridge ( In the vertical direction, scan the binary mask of the ridge to the left and right; (3) Location: When the pixel value changes from the body (e.g., 1) to the background (0), it is considered that the left boundary of the point has been found. ) and right boundary ( By iterating through all sampling points, a series of paired left and right boundary points can be obtained. (4) Output: These ordered pairs of boundary points ( , This constitutes the coordinate sequence of each ridge boundary.

[0133] In this embodiment, the average width of the local ridges of all ridges is calculated to obtain the average ridge width of the farmland:

[0134] Where N is the number of ridges, and M is the number of sampling points on the centerline of a single ridge. and These are the left and right boundary pixel coordinates of the i-th sampling point of the k-th ridge, respectively.

[0135] In this embodiment, a closed polygonal region is generated as the ridge region based on the coordinate sequence of the ridge boundary: The coordinate sequence of the ridge boundary is converted into standardized JSON format data; a closed polygonal region is generated based on the standardized JSON format data to obtain the ridge region (i.e., the final simplified ROI), providing a spatial reference for subsequent applications such as seedling monitoring and agricultural machinery navigation.

[0136] More specifically: The boundary coordinate sequence of each ridge is encapsulated into a list structure. Then, using a module such as Python's json, the list is serialized into a standardized JSON string. The JSON object can contain fields such as identifiers and coordinate pair sequences to ensure the structured and self-descriptive nature of the data. Finally, based on the coordinate sequence in JSON format, the closed polygon region is reconstructed in the computing environment. This can be done by reading the coordinate points in the JSON using a computational geometry library (such as OpenCV's fillPoly function or Shapely library) and connecting the first and last points in sequence to directly generate a filled binary mask, or by constructing a polygonal geometry object in a vector graphics system. This closed polygon region is the ridge region.

[0137] In this embodiment, the effective seedling sequence is constructed as follows in the seedling identification step: Convert the pixel coordinates of the position of each seedling in the ridge area to physical coordinates; Based on the ridge reference angle, the physical coordinates of the position of each seedling in each row of the ridge are projected and sorted along the extension direction of the ridge to obtain the effective seedling sequence.

[0138] More specifically: (1) Coordinate system transformation: Convert the pixel coordinates of the seedlings in the image to the physical coordinates in the field.

[0139] Establish a mapping relationship: using the calibration parameters of the farmland image (plot image) acquisition equipment (e.g., camera focal length, shooting height, etc.) and the pre-calculated average ridge width ( Determine the pixel-to-physical scale conversion coefficients. (Unit: meters per pixel). This coefficient establishes the proportional relationship between the pixel distance in the image and the actual physical distance.

[0140] Coordinate transformation calculation: using pixel-to-physical scale transformation coefficients It can identify the pixel coordinates of each seedling by a deep learning model (such as the YOLOv11n model). Convert to corresponding physical coordinates ,like ; (2) Projection and sorting: After obtaining the physical coordinates, the seedlings are sorted according to the direction of the ridges to form an effective sequence.

[0141] Determine the projection reference: The key parameter is the ridge reference angle ( ), which defines the direction of ridge extension.

[0142] Projection and sorting: Project the physical coordinates of each seedling onto a center line of the ridge that is parallel to the reference angle of the ridge and as close as possible to these points; then, sort the projected points in ascending or descending order according to the coordinate values ​​of these projected points along the direction of the ridge (e.g., from one end of the field to the other); this results in an ordered "effective seedling sequence" that reflects the true order of the seedlings on the ridge.

[0143] Implementation method 5: The method further includes: The plot analysis steps are as follows: Based on the location of each seedling in the obtained ridge area, the actual number of seedlings, and the effective seedling sequence, the sowing quality in the ridge area is evaluated by judging uniformity, deviation, evenness, and missing seedlings and broken rows. in: Uniformity: The uniformity of seedling distribution on the ridge is reflected by the uniformity of spacing and the uniformity of position; the uniformity of spacing is used to reflect the consistency of the distance between adjacent seedlings; the uniformity of position is used to reflect the degree of concentration of seedlings on the center line of the ridge. Degree of deviation: indicates the degree of lateral deviation of the growth position of a single seedling relative to the center line of the ridge; Uniformity: Indicates the consistency of the spacing between adjacent seedlings on the ridge, reflecting the uniformity of planting density; Missing seedlings: This indicates that one or more seedlings are missing; Broken rows: This indicates that multiple seedlings are missing from a row, forming a long blank section.

[0144] In this embodiment, uniformity refers to the evenness of the distribution of seedlings on the ridge, which is a comprehensive indicator. It is measured by two aspects: spacing uniformity (reflecting the consistency of the distance between adjacent seedlings) and position uniformity (reflecting the degree of concentration of seedlings on the center line of the ridge).

[0145] Spacing uniformity Based on the mean of adjacent physical distances between seedlings and standard deviation calculate: ; Spacing uniformity The value range is [0,1]. The closer the value is to 1, the more uniform the spacing.

[0146] Positional neatness Based on the mean vertical distance from the seedling to the center line of the ridge ( ) and half the average ridge width ( )calculate: ; Positional neatness The value range is [0,1]. The closer the value is to 1, the closer the position is to the center line.

[0147] Overall neatness S: The weights (such as 0.6, 0.4) can be adjusted.

[0148] Based on the overall uniformity S, the sowing quality in the ridge area can be graded, such as: S≥0.85 is “Excellent”; 0.7≤S<0.85 is “Good”; 0.55≤S<0.7 is “Average”; S<0.55 is “Poor”.

[0149] In this embodiment, the degree of deviation indicates the lateral deviation of the growth position of a single seedling relative to the centerline of the ridge. Specifically, it measures the uprightness of the seedling stem or whether there is a lateral deviation exceeding the allowable threshold, reflecting the "centering" or "uprightness" of sowing or transplanting.

[0150] Calculate the value of each seedling along the direction perpendicular to the reference angle of the ridge. Physical vertical distance to the center line of the ridge Statistics of all maximum value mean and over-skew ratio (Right now The proportion of seedlings, of which (Maximum allowable deviation threshold): when and At that time, it was "no card deviation"; when and At that time, it was described as "slight deviation"; when and At that time, it was classified as "moderate cardioversion"; when or At that time, it was described as "severe card deviation".

[0151] In this embodiment, uniformity refers to the consistency of the spacing between adjacent seedlings on the ridge, reflecting the uniformity of planting density.

[0152] Calculate the adjacent physical distance of an ordered seedling sequence Mean of spacing and standard deviation Thus, the coefficient of variation ( The smaller the coefficient of variation, the more uniform the seedling distribution. when At that time, it is considered "extremely uniform"; when At that time, it was described as "relatively uniform"; when At that time, it was considered "generally uniform"; when When the time is right, it is considered "uneven".

[0153] In this embodiment, the determination of missing seedlings and gaps in rows is based on determining the normal seedling spacing according to the crop variety and planting density. Set a threshold for determining missing seedlings ( ), threshold for determining the length of a broken row ( ); Number of missing seedlings calculate( (This is a rounding function). Criteria for determining a broken row: If the spacing between adjacent m consecutive seedlings meets the condition... And the cumulative length satisfies If the condition is met, the area is determined to be a broken row, and the length of the broken row is the cumulative length.

[0154] Implementation method 6: The method further includes: Field analysis steps: Based on the location of each seedling in the obtained ridge area, the actual number of seedlings, and the effective seedling sequence, calculate the seedling planting density of each plot and evaluate the straightness of the ridge.

[0155] In this embodiment, the seedling planting density for each plot is calculated: Based on the seedling detection results (position of each seedling in the ridge region, actual number of seedlings, and effective seedling sequence) output by the YOLOv11n model, the number of effective seedlings in each cell is counted; combined with the pixel-to-physical scale conversion coefficient. Calculate the actual planting density of seedlings (plants / square meter) for each plot: ; in, Let be the number of seedlings in the k-th cell. Let be the actual area of ​​the kth cell.

[0156] In this embodiment, the straightness of the ridge is evaluated. Based on the ridge-direction baseline extracted by Hough transform, the vertical deviation distance between the center line of each ridge and the ridge-direction baseline is calculated. The average deviation distance of a single ridge is statistically analyzed. and maximum deviation distance It represents the level of straightness.

[0157] Straightness evaluation indicators: ; in, The maximum allowable deviation distance according to the standard; Straightness evaluation index The value range is [0,1], and the closer the value is to 1, the better the straightness.

[0158] In this embodiment, by combining the analysis of each plot in the entire field with the preliminary seedling shortage diagnosis data of abnormal areas, a comprehensive report on seedling condition detection can be output, which may include, for example, basic information of the plot and detection results of core indicators (quantity, location coordinates, grade).

[0159] At the same time, based on the detected indicators, corresponding rectification suggestions can be given (such as density adjustment, straightness optimization, and replanting of missing seedlings, etc.) to achieve comprehensive and precise management of field seedling conditions.

[0160] Implementation Method 7: A field crop seedling condition diagnosis device based on UAV remote sensing images, the device comprising the following modules: Data acquisition module: acquires multiple plot images of fields awaiting seedling condition diagnosis based on UAV remote sensing technology; performs preliminary identification of abnormal areas in each plot image; for the preliminarily identified abnormal areas, acquires a point map of the abnormal area based on UAV remote sensing technology, and obtains preliminary seedling shortage diagnosis data for the abnormal area; wherein, the resolution of the point map of the abnormal area is higher than that of the plot image. Cellular ridge identification module: A pre-trained DeepLabV3+ model is used as the ridge segmentation model; each cell image is input into the ridge segmentation model to obtain a binary mask of the ridge, where each pixel is labeled as either the ridge or the background; a geometric processing algorithm is used to filter out the background from the binary mask of the ridge, retaining the ridge pixel region, and calculating the ridge boundary coordinate sequence; based on the ridge boundary coordinate sequence, a closed polygon region is generated as the ridge region, and an image of the identified ridge region is obtained; The seedling identification module on the ridge: The pre-trained YOLOv11n model is used as the seedling detection model on the ridge; the image of the identified ridge area is input into the seedling detection model on the ridge to detect individual seedlings, obtain the position of each seedling in the ridge area and the actual number of seedlings, and construct an effective seedling sequence.

[0161] Implementation Method 8: A computer device comprising: a processor and a memory, the memory for storing executable instructions of the processor, the processor being configured to execute the field crop seedling condition diagnosis method based on UAV remote sensing images according to any one of the above embodiments by executing the executable instructions.

[0162] Implementation Method 9: A computer storage medium storing a computer program, wherein when the computer program is executed, the field crop seedling condition diagnosis method based on UAV remote sensing images described in any one of the above implementation methods is performed.

[0163] Implementation Method 10: A computer program product, comprising a computer program / instructions, which, when executed by a processor, implements the steps of the field crop seedling condition diagnosis method based on UAV remote sensing images described in any of the above implementation methods.

[0164] Implementation Method 8: A specific implementation method is provided: Application scenario: Seedling condition diagnosis in a 200-mu soybean field at the 2-leaf stage, with soybean seedlings 5-8cm tall and 3-5cm wide, planted in horizontal ridges.

[0165] Implementation steps: (1) Drone field inspection: DJI M300 RTK drones are used, equipped with high-definition RGB cameras and DJI embedded devices and detection equipment. The flight altitude is 15 meters and the speed is 5 m / s, so that the drone can detect the ridge direction and fly at the same height along the ridge direction. One picture covers a 20m×20m plot with no repetition.

[0166] (2) Fixed-point flight: Based on the DJI detection equipment and the canopy coverage inversion results, four abnormal areas were identified and marked. The drone was then flown to the abnormal areas again to collect high-resolution RGB images at a height of 5m along the ridge direction. The initial detection results were output simultaneously, and four areas with missing seedlings were identified. The results of these four areas were then subjected to focused detection.

[0167] (3) Ridge identification: The DeepLabV3+ semantic segmentation model is used to segment the ridge region of a single cell and output the binary mask of the ridge in each cell. Based on the binary mask of the ridge, the ROI region in each cell is accurately delineated (only the ridge region is retained, and the background areas such as the gaps between ridges and the field ridges are filtered out) to limit the range for subsequent seedling detection.

[0168] (4) Seedling detection: The optimized YOLOv11n model was used to detect seedlings in the divided ridge area; the confidence threshold was set to 0.4 (to reduce the false detection rate of weeds) to take into account the small crown width (3-5cm) of soybeans at the 2-leaf stage; high-confidence seedling targets were detected in the ridge area, and false detection targets outside the ridge were filtered out. The model detection accuracy reached more than 95%, and the number and coordinates of seedlings in each plot were successfully identified.

[0169] (6) Seedling condition analysis: The average density of soybean seedlings was 90 plants / m². 2 Anomaly detection: Analyze areas initially identified as having missing seedlings, calculating the area of ​​the missing seedlings to be 2-3 square meters. 2 All were located within the ridge area, 3 seedling deviation areas (a total of 15 seedlings deviated from the ridge centerline by more than 12.5cm), and 5 areas of abnormal uniformity (coefficient of variation CV_d=0.32-0.45); comprehensive judgment: the spacing uniformity and position uniformity were calculated, and the overall uniformity was calculated at the same time. The average S=0.78 for the whole field, and the seedling condition was judged as "normal".

[0170] (7) Develop targeted measures based on the diagnostic report: Replanting in areas with missing seedlings: Plant 90 seedlings / m² in 4 areas with missing seedlings. 2 Replant two-leaf stage soybean seedlings from the same batch, and irrigate and moisturize them promptly after replanting; adjust the offset areas: manually straighten the seedlings in the three offset areas to ensure that they do not deviate from the ridge during subsequent growth.

[0171] The above description of several specific embodiments further details the technical solution provided by the present invention in order to highlight the advantages and benefits of the technical solution provided by the present invention. However, the above-described specific embodiments are not intended to limit the present invention. Any reasonable modifications and improvements to the present invention, reasonable combinations of embodiments, and equivalent substitutions based on the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for diagnosing crop seedling conditions in field crops based on UAV remote sensing images, characterized in that, The method includes the following steps: Data acquisition steps: acquire multiple plot images of fields awaiting seedling condition diagnosis based on UAV remote sensing technology; preliminarily identify abnormal areas in each plot image; for the preliminarily identified abnormal areas, acquire point maps of the abnormal areas based on UAV remote sensing technology, and obtain preliminary seedling shortage diagnosis data for the abnormal areas; wherein, the resolution of the point maps of the abnormal areas is higher than that of the plot images. Cellular ridge identification steps: A pre-trained DeepLabV3+ model is used as the ridge segmentation model; each cell image is input into the ridge segmentation model to obtain a binary mask of the ridge body, where each pixel is labeled as either the ridge body or the background; a geometric processing algorithm is used to filter out the background from the binary mask of the ridge body, retaining the ridge body pixel region, and calculating the ridge body boundary coordinate sequence; based on the ridge body boundary coordinate sequence, a closed polygonal region is generated as the ridge body region, and an image of the identified ridge body region is obtained; The seedling identification steps on the ridge are as follows: The pre-trained YOLOv11n model is used as the seedling detection model on the ridge; the image of the identified ridge area is input into the seedling detection model on the ridge, and the individual seedlings are detected to obtain the position of each seedling in the ridge area and the actual number of seedlings, and to construct an effective seedling sequence.

2. The method for diagnosing crop seedling conditions in field crops based on UAV remote sensing images according to claim 1, characterized in that, In the data acquisition step, the cell image is acquired using the following method: Select fields at the emergence stage based on the target crop type as fields for seedling condition diagnosis; Drones were used to patrol the fields for crop condition diagnosis in order to collect images of the fields and the crops on them, and to obtain multiple images of the plots. During field patrol operations: The flight path of the drone is planned parallel to the direction of the field where the crop condition needs to be diagnosed: the drone flies at a constant speed along the direction of the ridge to collect images, ensuring that the images collected by each drone are not repeated and that all images collected by the drones cover the entire area of ​​the field where the crop condition needs to be diagnosed; the actual field area corresponding to each collected image is equal, and the actual field area corresponding to a single collected image is defined as a small area. The drone collects data at the same altitude relative to the ground.

3. The method for diagnosing crop seedling conditions in field crops based on UAV remote sensing images according to claim 1, characterized in that, In the data acquisition step, abnormal areas in each cell image are initially identified; for the initially identified abnormal areas, a point map of the abnormal areas is acquired based on UAV remote sensing technology, and preliminary seedling shortage diagnosis data for the abnormal areas is obtained; wherein, the resolution of the point map of the abnormal areas is higher than that of the cell images, including the following: For each cell image, vegetation index and canopy coverage are detected. Areas with vegetation index and canopy coverage below a given vegetation threshold are marked and pinpointed according to their coordinates to delineate abnormal areas. The anomaly region mapping was collected using the following method: The drone flew directly above all the marked points and began acquiring images of the abnormal area; The drone flies at a constant speed along the direction of the ridge to collect images of abnormal areas; the direction of the ridge extension when collecting images of abnormal areas is the same as the direction of the ridge extension when collecting images of the plot. The drone's acquisition altitude relative to the ground is the same, and the acquisition altitude when acquiring images of abnormal areas is lower than the acquisition altitude when acquiring images of the cell. The seedling emergence status is inverted by using the collected point maps of abnormal areas to obtain preliminary seedling loss diagnosis data for abnormal areas.

4. The method for diagnosing crop seedling conditions in field crops based on UAV remote sensing images according to claim 1, characterized in that, In the ridge identification step, a geometric processing algorithm is used to filter out the background from the ridge binary mask, retain the ridge pixel region, and calculate the ridge boundary coordinate sequence, including the following steps: Steps for extracting the centerline of a ridge: Obtain the probability map of the binary mask of the ridge; assuming that the ridges are arranged in rows and columns in the probability map, and the image width direction is taken as the column direction; where the predicted probability of the ridge in the k-th row and j-th column is... , ; Calculation along the column direction satisfies The weighted average of the pixel column coordinates is used as the column coordinates of the centerline of the ridge to obtain the centerline of the ridge: ； in, represents the column coordinates of the center line of the k-th row of the ridge; W is the image width of the probability plot, in pixels; This is an indicator function; I = 1 when the condition inside the parentheses is true, and I = 0 otherwise. Segmentation threshold, ; Row spacing calculation steps: Based on the centerline column coordinates of two adjacent rows of rows, calculate the spacing between the two adjacent rows of rows, and take the average spacing as the pixel-level average row spacing. ； Where d is the average row spacing; N is the number of rows; Scale conversion steps: Convert pixel-level average row spacing to physical-scale row spacing: ； in, Pixel-to-physical scale conversion coefficient; The steps for extracting the ridge direction baseline are as follows: In the pixel coordinate system, the coordinates of the center line column of the obtained ridge are detected. The Hough transform in polar coordinate form is used to transform the detection of the ridge direction baseline into peak detection in the parameter space. The angle distribution of the candidate ridge direction baselines obtained by detection is statistically analyzed. After removing outliers, the average of the high-frequency angles is taken as the ridge direction baseline angle. Based on the ridge direction baseline angle, three types of ridge direction are divided: horizontal ridge direction, vertical ridge direction, and oblique ridge direction. The steps for calculating ridge width are as follows: A series of uniformly distributed sampling points are set at certain intervals along the centerline of the ridge; at each sampling point, the binary mask of the ridge is scanned to the left and right sides along a direction perpendicular to the ridge reference angle; when a pixel value changes from the ridge to the background, it is determined as the left or right boundary point of the ridge; all sampling points are traversed to obtain a series of ordered boundary point pairs consisting of left and right boundary points, which serve as the ridge boundary coordinate sequence; the pixel distance between the left and right boundary points of each sampling point is calculated, and the average value is taken as the local ridge width; the average ridge width of all ridges is calculated by averaging the local ridge widths to obtain the average ridge width of the farmland.

5. The method for diagnosing crop seedling conditions in field crops based on UAV remote sensing images according to claim 1, characterized in that, The method further includes: The plot analysis steps are as follows: Based on the location of each seedling in the obtained ridge area, the actual number of seedlings, and the effective seedling sequence, the sowing quality in the ridge area is evaluated by judging uniformity, deviation, evenness, and missing seedlings and broken rows. in: Uniformity: The uniformity of seedling distribution on the ridge is reflected by the uniformity of spacing and the uniformity of position; the uniformity of spacing is used to reflect the consistency of the distance between adjacent seedlings; the uniformity of position is used to reflect the degree of concentration of seedlings on the center line of the ridge. Degree of deviation: indicates the degree of lateral deviation of the growth position of a single seedling relative to the center line of the ridge; Uniformity: Indicates the consistency of the spacing between adjacent seedlings on the ridge, reflecting the uniformity of planting density; Missing seedlings: This indicates that one or more seedlings are missing; Broken rows: This indicates that multiple seedlings are missing from a row, forming a long blank section.

6. The method for diagnosing crop seedling conditions in field crops based on UAV remote sensing images according to claim 1, characterized in that, The method further includes: Field analysis steps: Based on the location of each seedling in the obtained ridge area, the actual number of seedlings, and the effective seedling sequence, calculate the seedling planting density of each plot and evaluate the straightness of the ridge.

7. A field crop seedling condition diagnosis device based on UAV remote sensing images, characterized in that, The device includes the following modules: Data acquisition module: acquires multiple plot images of fields awaiting crop condition diagnosis based on UAV remote sensing technology; performs preliminary identification of abnormal areas in each plot image; For the initially identified abnormal areas, a point map of the abnormal areas was collected based on UAV remote sensing technology, and preliminary seedling shortage diagnosis data for the abnormal areas was obtained; among them, the resolution of the point map of the abnormal areas is higher than that of the cell image. Cellular ridge identification module: A pre-trained DeepLabV3+ model is used as the ridge segmentation model; each cell image is input into the ridge segmentation model to obtain a binary mask of the ridge, where each pixel is labeled as either the ridge or the background; a geometric processing algorithm is used to filter out the background from the binary mask of the ridge, retaining the ridge pixel region, and calculating the ridge boundary coordinate sequence; based on the ridge boundary coordinate sequence, a closed polygon region is generated as the ridge region, and an image of the identified ridge region is obtained; The seedling identification module on the ridge: The pre-trained YOLOv11n model is used as the seedling detection model on the ridge; the image of the identified ridge area is input into the seedling detection model on the ridge to detect individual seedlings, obtain the position of each seedling in the ridge area and the actual number of seedlings, and construct an effective seedling sequence.

8. A computer device, comprising: The processor and memory are characterized in that the memory is used to store executable instructions of the processor, the processor being configured to execute the field crop seedling diagnosis method based on UAV remote sensing images according to any one of claims 1-6 by executing the executable instructions.

9. A computer storage medium, characterized in that, The storage medium stores a computer program, which, when executed, performs the field crop seedling condition diagnosis method based on UAV remote sensing images as described in any one of claims 1-6.

10. A computer program product comprising a computer program / instructions, characterized in that, When the computer program / instruction is executed by the processor, it implements the steps of the field crop seedling condition diagnosis method based on UAV remote sensing images as described in any one of claims 1-6.

Images