Method and system for measuring number (area) of plants in tobacco field based on visible light image of unmanned aerial vehicle

Abstract

The invention relates to the technical field of agricultural remote sensing monitoring, in particular to a method and system for measuring the number (area) of plants in a tobacco field based on visible light images of an unmanned aerial vehicle, and is suitable for automatic and high-precision growth monitoring and resource accounting of a large-scale tobacco field. Comprising six steps of tobacco field visible light image acquisition, image preprocessing, tobacco field region segmentation, tobacco field plant target detection and plant number statistics, tobacco field area calculation, and result post-processing and output, and through combination of an improved target detection algorithm and an image segmentation technology, high precision of tobacco field plant number statistics and high accuracy of area determination are realized. Meanwhile, the detection efficiency is guaranteed, the actual requirement of large-scale tobacco field monitoring is met, data acquisition, processing, analysis and output can be automatically completed in the whole process, manual intervention is not needed, the monitoring efficiency is greatly improved, and the labor cost is reduced.

Description

Methods and Systems for Determining Tobacco Plant Count (Area) Based on UAV Visible Light Imagery Technical Field

[0001] This invention relates to the field of agricultural remote sensing monitoring technology, specifically to a method and system for determining the number and area of ​​tobacco plants in a field based on visible light images from unmanned aerial vehicles (UAVs), which is applicable to automated, high-precision growth monitoring and resource accounting in large-scale tobacco fields. Background Technology

[0002] The counting of tobacco plants and the determination of the area are core foundational data for tobacco planting management, yield forecasting, and resource allocation. Traditional methods rely on manual on-site counting and measurement, which suffers from low efficiency, high labor intensity, strong subjectivity, and difficulty in covering large-scale tobacco fields. Furthermore, they are easily affected by terrain and crop growth stage, resulting in insufficient data accuracy and timeliness.

[0003] Existing image-based monitoring technologies partially utilize satellite remote sensing or fixed ground cameras for image acquisition. However, satellite remote sensing has limited resolution, making it difficult to identify individual tobacco plants. Furthermore, ground-based acquisition is limited by viewing angle, resulting in small coverage areas and cumbersome operation. While UAV remote sensing technology, with its advantages of flexibility, efficiency, and high resolution, is increasingly being applied to agricultural monitoring, it still faces key technical bottlenecks in tobacco field scenarios: tobacco plants are easily obscured by leaves in visible light images, and dense row spacing, background interference (soil, weeds), along with the high false negative rate, severe false positives, and insufficient accuracy in plant counts of traditional target detection algorithms (such as Faster-RCNN and SSD). Additionally, tobacco field boundaries are often irregular, making it difficult for traditional image segmentation methods to accurately distinguish between tobacco and non-tobacco fields, leading to significant errors in area measurement.

[0004] Existing UAV target detection technologies have limited ability to extract small target features in complex scenarios, exhibit spatial and channel feature redundancy, and lack robustness in identifying densely packed, occluded tobacco plants. Therefore, there is an urgent need for a technical solution that can accurately extract small target features of tobacco plants and efficiently segment tobacco field areas, enabling automated and high-precision measurement of the number and area of ​​tobacco plants. Summary of the Invention

[0005] This invention provides a method and system for determining the number (area) of tobacco plants in a field based on visible light images from a UAV. By combining an improved target detection algorithm with image segmentation technology, it achieves high precision in tobacco plant count and high accuracy in area measurement, while ensuring detection efficiency and meeting the actual needs of large-scale tobacco field monitoring.

[0006] To achieve the above objectives, the technical solution adopted by this invention is as follows: A method for determining the number (area) of tobacco plants in a field based on visible light images from a UAV, comprising the following steps: S1. Acquisition of visible light images of tobacco fields: A UAV equipped with a visible light image sensor is operated to take full-coverage photos of the target tobacco field according to a preset flight path, acquiring visible light images and recording UAV attitude parameters, including flight altitude, latitude and longitude, and heading angle, during the acquisition process for subsequent image georegistration; S2. Image preprocessing: The acquired visible light images are preprocessed, and image georegistration is performed based on the UAV attitude parameters to eliminate image shifts caused by flight attitude fluctuations; the dataset is expanded using an improved Mosaic data augmentation method, while brightness equalization, noise removal, and image cropping are performed, with a unified output size of 640×640 pixels to improve model robustness; S3. Tobacco field region segmentation: An improved semantic segmentation network is used to extract tobacco field regions from the preprocessed images. Using U-Net as the base network, an SCConv module is embedded at the encoding end to reduce feature redundancy in the background region and enhance the feature representation of the tobacco field region; simultaneously, a CBAM convolutional block attention mechanism is introduced at the decoding end to strengthen the tobacco field region segmentation. S4. Tobacco Field Plant Target Detection and Counting: Based on the segmented tobacco field regions, an improved YOLO-SCAT algorithm is used for small target detection of tobacco plants. Using YOLOv7 as the baseline network, the last ELAN module of the backbone feature extraction network is replaced with an ELAN-SC module, and a CBAM attention mechanism is introduced after the ELAN-SC module. The detection head outputs the predicted bounding box, confidence score, and category information of the tobacco plants, and a confidence threshold (≥0.5) is set. S5. Calculate the area of ​​the tobacco field: Combine the tobacco field segmentation mask and the geographical parameters collected by the UAV to calculate the area. Based on the UAV flight altitude and image sensor parameters, calculate the image pixel resolution. Count the total number of foreground pixels in the tobacco field segmentation mask and calculate the actual area of ​​the tobacco field based on the pixel resolution. S6. Post-processing and output of results: Post-process the plant count and area calculation results, remove abnormal detection targets in the edge areas, and output the plant count and area data and visualization report for each region based on the distribution of tobacco field plots.

[0007] Further, step S1 specifically includes: S1.1 Parameter calculation and equipment calibration: The UAV attitude data heading angle, roll angle and pitch angle are processed by the image geographic coordinate transformation formula to calibrate the shooting angle; S1.2 Image resolution calculation: The UAV flight altitude H determines the ground sampling distance of the image.

[0008] Further, the specific steps of improving the Mosaic data augmentation method in step S2 are as follows: (1) Randomly select N tobacco field images, scale and flip them respectively, with a scaling ratio of 0.5-1.5 times and a horizontal / vertical flip probability of 50%; (2) Normalize the color space of each image to eliminate the influence of lighting differences, and stitch the N normalized images into one 640×640 pixel image. The stitching area adopts linear fusion transition and is stitched in groups of N=4.

[0009] Furthermore, in step S3, the process of embedding the SCConv module at the encoding end based on the U-Net network is divided into two parts: the Spatial Reconstruction Unit (SRU) is used to separate farmland plants from background features, and the Channel Reconstruction Unit (CRU) is used to reduce the redundancy of similar leaf texture channels.

[0010] Furthermore, in step S4, the CBAM attention mechanism is introduced after the ELAN-SC module, which is a channel attention mechanism and a spatial attention mechanism, to highlight the key channels of the plant and the spatial position of the plant.

[0011] Furthermore, in step S5, the actual area of ​​the tobacco field is related to the number of foreground pixels and the pixel resolution. When the UAV has a roll angle β or a pitch angle θ, the image will produce perspective distortion, and the area needs to be corrected to finally obtain the actual area of ​​the tobacco field.

[0012] Furthermore, in step S6, the specific process of removing abnormal detection targets in the edge region is as follows: an "isolated detection box" is set up at the boundary of the tobacco field, and spatial clustering algorithm is used to remove them.

[0013] Furthermore, the specific process of the spatial clustering algorithm for elimination is as follows: extract the center coordinates of all detection boxes, and determine the pixel coordinate range of the paddy field area based on the paddy field segmentation mask; calculate the minimum distance from the center of each detection box to the paddy field area; set a distance threshold, and if the minimum distance is greater than the distance threshold, then the detection box is eliminated.

[0014] Furthermore, a system including a method for determining the number (area) of tobacco plants in a field based on visible light images from a UAV includes: an image acquisition module, a data preprocessing module, a tobacco field segmentation module, a target detection and counting module, an area calculation module, and a result output module. The image acquisition module consists of a UAV and a visible light image sensor for image acquisition. The data preprocessing module receives image data from the image acquisition module and performs preprocessing. The tobacco field segmentation module performs precise segmentation of the tobacco field area and background area from the preprocessed image data and outputs a tobacco field segmentation mask. The target detection and counting module performs small target detection of tobacco plants in the segmented tobacco field area and outputs the total number of tobacco plants. The area calculation module calculates the actual area of ​​the tobacco field based on the image pixel resolution, the number of foreground pixels in the tobacco field segmentation mask, and the geographical parameters of the UAV. The result output module performs post-processing on the plant count and area data to generate a visualization report.

[0015] Furthermore, the tobacco field segmentation module incorporates an improved U-Net semantic segmentation network, and the target detection and counting module deploys an improved YOLO-SCAT algorithm model through the SCConv module and the CBAM attention mechanism.

[0016] Compared with the prior art, the present invention has the following beneficial effects: (1) The present invention uses the improved YOLO-SCAT algorithm to detect tobacco plants in the field. The ELAN-SC module reduces feature redundancy and the CBAM attention mechanism strengthens small target features. The precision and recall rate can reach 94.4%, which greatly reduces the false negative and false positive rates. The plant count accuracy is higher than that of the existing algorithm.

[0017] (2) The tobacco field segmentation module of this invention combines SCConv and CBAM technologies to accurately extract the boundaries of tobacco fields with an area measurement error of ≤3%, which meets the accuracy requirements of agricultural production. Furthermore, the algorithm detection frame rate reaches 76.9 frames / s, which supports real-time processing of video streams collected by drones and is adapted to the high-efficiency monitoring needs of large-scale tobacco fields.

[0018] (3) The present invention can automatically complete the data collection, processing, analysis and output process without human intervention, which greatly improves the monitoring efficiency and reduces labor costs. In addition, the UAV system has strong compatibility and can be adapted to different models of UAVs and tobacco field scenarios, such as plains and mountain tobacco fields, with excellent generalization ability. Attached Figure Description

[0019] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used together with the embodiments of the invention to explain the invention and do not constitute a limitation thereof.

[0020] Figure 1 is a schematic diagram of the system structure for determining the number of tobacco plants (area) in the present invention; Figure 2 is a schematic diagram of the improved YOLO-SCAT algorithm model structure of the present invention; Figure 3 is a schematic diagram of the improved U-Net tobacco field segmentation network structure of the present invention; Figure 4 is a schematic diagram of the method flow for determining the number of tobacco plants (area) in the present invention. Detailed Implementation

[0021] The present invention will now be described in detail with reference to specific embodiments. These embodiments will help those skilled in the art to further understand the present invention, but do not limit the invention in any way. It should be noted that those skilled in the art can make several changes and improvements without departing from the concept of the present invention. These all fall within the protection scope of the present invention.

[0022] Please refer to Figures 1-4. This invention provides a method for determining the number (area) of tobacco plants in a field based on visible light images from a UAV, comprising the following steps: S1. Visible light image acquisition of the tobacco field: A UAV equipped with a visible light image sensor is operated to perform full-coverage photography of the target tobacco field according to a preset flight path (heading overlap ≥80%, lateral overlap ≥60%). The acquired image resolution is no less than 1920×1080 pixels, and the frame rate is 30Hz. Full-coverage photography of the target tobacco field is performed, acquiring visible light images and recording UAV attitude parameters during the acquisition process, including flight altitude, latitude and longitude, and heading angle, for subsequent image georegistration.

[0023] S1.1 Parameter Calculation and Equipment Calibration: To ensure the accuracy of image geographic coordinates, the UAV attitude data (heading angle, roll angle, and pitch angle) are processed using image geographic coordinate transformation formulas to calibrate the shooting angle; Let the UAV GPS coordinates be (X... U Y U Z U (Geodetic coordinate system), if the image pixel coordinates are (μ, v), then the corresponding geodetic coordinates (X) are (μ, v). G Y G The calculation is as follows: ;u 0， v0: Image center pixel coordinates (e.g., in a 1920×1080 image, u0=960, v0=540); pixel_size: Physical pixel size of the image sensor (e.g., for a 1 / 3-inch Type CMOS sensor, pixel-size=3.75μm); α: UAV heading angle (angle with true north), β: Roll angle (angle of the fuselage tilting left and right); Formula function: Establishes the mapping between image pixels and actual geographical location, laying the coordinate foundation for subsequent area calculation.

[0024] S1.2 Image Resolution Calculation (Associated with Flight Altitude): The UAV's flight altitude H (height above the ground, unit: m) determines the image ground sampling distance (GSD, i.e., the actual ground distance represented by each pixel), as shown in the following formula: Where f: the focal length of the visible light lens.

[0025] S2. Image preprocessing: The acquired visible light images are preprocessed, including image georegistration based on UAV attitude parameters to eliminate image offset caused by flight attitude fluctuations; the dataset is expanded using an improved Mosaic data augmentation method, while brightness equalization, noise removal and image cropping are performed, and the output size is uniformly set to 640×640 pixels to improve model robustness.

[0026] The specific process of improving the Mosaic data augmentation algorithm is as follows: (1) Randomly select N tobacco field images and perform random scaling (scaling ratio 0.5-1.5 times) and random flipping (horizontal / vertical flipping probability 50%) respectively; (2) Normalize the color space of each image to eliminate the influence of illumination differences. The formula is: I(u,v): Original image pixel value, μ RGB Training set images, σ RGB Training set images; RGB: three-channel standard deviation; N normalized images are stitched together into a single 640×640 pixel image. The stitching region uses a linear blending transition, typically grouped into N=4 groups, using the following formula: Among them, the α fusion weight (0≤α≤1) changes linearly with the pixel position to avoid harsh stitching edges.

[0027] S3. Tobacco Field Region Segmentation: An improved semantic segmentation network is used to extract the tobacco field region from the preprocessed image. U-Net is used as the base network, and the SCConv module is embedded in the encoder to reduce feature redundancy in the background region and enhance the feature representation of the tobacco field region. At the same time, the CBAM convolutional block attention mechanism is introduced in the decoder to strengthen the extraction of tobacco field boundary features, accurately distinguish between tobacco field and non-tobacco field regions, and output a binarized tobacco field segmentation mask.

[0028] In the U-Net encoder, the SCoConv module (spatial + channel modeling) is embedded. The core formula is divided into two parts: (1) Spatial Reconstruction Unit (SRU): Separates farmland vegetation from background features. The formula is: γ: Trainable parameter of group normalization (GN) layer, reflecting the richness of feature information; W1: Information-rich feature weight (plant features), W2: Redundant feature weight (soil / weeds); ⊗: Element-wise multiplication, ⊕: Feature concatenation; These functions enhance plant features and suppress background redundancy.

[0029] (2) Channel Reconstruction Unit (CRU): Reduces redundancy in channels of similar leaf textures, the formula is: (Divided according to channel ratio α=0.5, i.e. X) up (50% of the channels) , , GWC: Grouped depthwise convolution (capturing plant edge features), PWC: Pointwise convolution (compressing channel dimensions); S1, S 2: The global average pooling results of Y1 and Y2, β1 and β2: channel attention weights; similar leaf texture channels are merged to reduce computation while retaining key features.

[0030] S4. Tobacco Field Plant Target Detection and Plant Counting: Based on the segmented tobacco field area, an improved YOLO-SCAT algorithm is used for small target detection of tobacco field plants. Using YOLOv7 as the baseline network, the last ELAN module of the backbone feature extraction network is replaced with an ELAN-SC module (an ELAN structure reconstructed from SCConv), reducing spatial and channel redundancy features and lowering computational complexity. A CBAM attention mechanism is introduced after the ELAN-SC module to enhance the learning of small target features of tobacco field plants from both spatial and channel dimensions, suppressing background interference. The detection head outputs the predicted bounding box, confidence level, and category information of tobacco field plants. A confidence threshold (≥0.5) is set to filter false detection targets. The detected tobacco field plants are counted to obtain the total number of tobacco field plants.

[0031] (1) The ELAN-SC module configuration process involves replacing the last ELAN module in the YOLOv7 backbone network with the ELAN-SC module. This layer introduces SCoConv to reduce tobacco planting redundancy features; the module output feature calculation formula is as follows: ;F branch1 Branch 1 output (1×1 convolution + BN + SiLU, preserving global information); F branch2 Branch 2 input ((3×3 convolution)×2), processed by SCConv to obtain SCConv(F) branch2 (Reduce redundant tobacco planting features); Concat: Feature concatenation to ensure that the output feature dimension is consistent with the original ELAN (e.g., if the number of input channels is 256, the output is still 256), ensuring network compatibility.

[0032] (2) Introducing the CBAM attention mechanism: After the ELAN-SC module, the CBAM attention mechanism is introduced to enhance the characteristics of tobacco plants in the field from the perspectives of channels and space. The formula is divided into two parts: 1) Channel attention mechanism: highlighting key channels of the plant (such as leaf edge channels), the formula is: AvgPool: Global Average Pooling; MaxPool: Global Max Pooling; MLP: Multilayer Perceptron (containing 1 hidden layer, channel compression ratio r=16); By pooling, channel information is aggregated, and MLP learns channel weights to enhance the key channel features for plant identification.

[0033] 2) Spatial attention mechanism: Emphasizing the plant's spatial location (e.g., the central area of ​​the plant), the formula is as follows: , : Feature map after channel attention weighting; f 7×7 7*7 convolution (captures large-scale spatial relationships), Concat: channel feature concatenation (concatenates two single-channel pooling images into a 2-channel image), focusing on the spatial region where the spider is located in the image and suppressing interference from background areas such as soil and weeds.

[0034] S5. Tobacco Field Area Calculation: Calculate the area by combining the tobacco field segmentation mask and the geographical parameters collected by the UAV. Calculate the image pixel resolution (actual geographical distance / pixel) based on the UAV flight altitude and image sensor parameters. Count the total number of foreground pixels in the tobacco field segmentation mask and calculate the actual area of ​​the tobacco field (Area = Number of foreground pixels * Pixel resolution²).

[0035] (1) Pixel statistics of tobacco fields Let the tobacco field segmentation mask be M(u,v), where M(u,v)=1 represents the tobacco field area and M(u,v)=0 represents the non-tobacco field area. Then the total number of pixels N in the tobacco field area is N. total The calculation formula is: W and H represent the image width and height, respectively.

[0036] (2) Area Correction: When the UAV has a roll angle β or a pitch angle θ, the image will produce perspective distortion, and the area needs to be corrected. The formula is: The basic area calculation is based on GSD, where the actual tobacco field area S is... base The formula is: .

[0037] S6. Post-processing and output of results: Post-process the plant count and area calculation results: Remove abnormal detection targets in the edge areas (such as isolated detection boxes outside the tobacco field boundary); Combine the distribution of tobacco field plots to output plant count and area data and visualization reports for each region (including tobacco field distribution map, plant count heat map, and area calculation table).

[0038] For setting "isolated detection boxes" (such as false detection targets in non-tobacco field areas) at the boundary of tobacco fields, spatial clustering algorithm is used to remove them. The steps are as follows: (1) Extract the center coordinates (x, y, x) of all detection boxes. c y c (and determine the pixel coordinate range of the paddy field area based on the paddy field segmentation mask). (2) Calculate the minimum distance from the center of each detection frame to the paddy field area, using the following formula: Other distances (d>0) indicate that the center of the detection box is outside the rice paddy area, and are isolated detection candidates; a distance threshold (d) is set. T =50) pixels (set according to the size of the rice paddy image), if (d(u c ,v c )>d T If the value is 0, then the detection box is removed.

[0039] The system for determining the number of tobacco plants (area) in a field based on visible light images from unmanned aerial vehicles includes: an image acquisition module, a data preprocessing module, a tobacco field segmentation module, a target detection and counting module, an area calculation module, and a result output module.

[0040] Image acquisition module: It consists of a drone platform and a visible light image sensor. The drone is equipped with a GPS / BeiDou positioning module and an attitude sensor to acquire visible light images of tobacco fields and flight attitude data. The visible light image sensor adopts a 1 / 3-inch Type CMOS sensor and supports image acquisition of 1920×1080 pixels at 30Hz.

[0041] Data preprocessing module: Used to receive raw data from the image acquisition module, perform brightness equalization, noise removal (Gaussian filtering), and image cropping, and output standardized preprocessed image data.

[0042] Tobacco Field Segmentation Module: It incorporates an improved U-Net semantic segmentation network and uses the SCConv module and CBAM attention mechanism to achieve accurate segmentation of the tobacco field region from the background region, and outputs the tobacco field segmentation mask.

[0043] Target detection and counting module: Deploys an improved YOLO-SCAT algorithm model to detect small targets of tobacco plants in the segmented tobacco field area, filters out false targets, completes the count of plants, and outputs the total number of tobacco plants in the field.

[0044] Area calculation module: Based on the image pixel resolution and the number of foreground pixels in the tobacco field segmentation mask, combined with UAV geographic parameters, the actual area of ​​the tobacco field is calculated.

[0045] Results output module: Used to post-process plant number and area data to generate visualization reports (including data tables, distribution maps, and heat maps), and supports data export (Excel and CSV formats) and real-time display.

[0046] (1) Experimental environment configuration: DJI M300RTK, equipped with a 1 / 3 Type CMOS visible light sensor; hardware environment: Intel Xeon Gold 5120 CPU, NVIDIA GeForce RTX 2080ti GPU, 32GB memory; software environment: Ubuntu 18.04 operating system, Python 3.10, PyTorch 1.12 deep learning framework; dataset: 800 visible light images of tobacco fields in different regions (plains, mountains) and different growth stages (seedling stage, vigorous growth stage) were collected, of which 640 images were used as the training set, 80 images as the validation set, and 80 images as the test set. The LabelMe tool was used to label the tobacco plant targets and tobacco field boundaries.

[0047] (2) Improved model training and parameter settings: YOLO-SCAT model training parameters: 300 training rounds, batch size=12, initial learning rate 0.01, momentum 0.937, decay coefficient 0.0005, SGD optimizer is used; Improved U-Net model training parameters: 200 training rounds, batch size=8, initial learning rate 0.001, Adam optimizer is used, loss function is Dice loss + cross-entropy loss.

[0048] (3) Implementation steps: Plan the UAV flight path: Based on the target tobacco field area, set the flight altitude to 100 meters, the forward overlap rate to 85%, and the lateral overlap rate to 65%, and start the UAV to collect full-coverage images; Data preprocessing: Use the data preprocessing module to perform georegistration (combined with GPS data), brightness equalization, and noise removal on the collected raw images, and crop them to 640×640 pixel images; Tobacco field segmentation: Input the preprocessed images into the tobacco field segmentation module, output a binarized tobacco field segmentation mask, and accurately remove background areas such as roads and weeds; Statistical Analysis: The tobacco field segmentation mask is input into the target detection and counting module. The improved YOLO-SCAT algorithm is used to detect tobacco plants. A confidence threshold of 0.5 is set to filter out false detection targets and count the total number of plants. Area Calculation: The area calculation module calculates the pixel resolution (0.05 m / pixel) based on the UAV's flight altitude (100 meters), counts the number of foreground pixels in the segmentation mask, and calculates the actual area of ​​the tobacco field. Result Output: The result output module generates a tobacco field plant count table, an area calculation table, and a visual distribution map. Data can be exported to Excel format.

[0049] (4) The results of the effectiveness verification test set show that the accuracy of tobacco plant detection in this system is 94.2%, recall rate is 94.0%, F1 score is 94.1%, plant count error is ≤2.5%, tobacco field area measurement error is ≤2.8%, and detection frame rate is 75.3 frames / s, which meets the requirements of high efficiency and high precision monitoring of large-scale tobacco fields.

[0050] Specific embodiments of the present invention have been described above. It should be understood that the present invention is not limited to the specific embodiments described above, and those skilled in the art can make various changes or modifications within the scope of the claims, which do not affect the essence of the present invention. Unless otherwise specified, the embodiments and features described in this application can be arbitrarily combined with each other.

Claims

1. A method for determining the number (area) of tobacco plants in a field based on visible light images from a UAV, characterized in that, Includes the following steps: S1. Visible light image acquisition of tobacco fields: Control the drone equipped with a visible light image sensor to take full-coverage pictures of the target tobacco field according to the preset flight path, acquire visible light images and record the drone's attitude parameters, including flight altitude, latitude and longitude, and heading angle, for subsequent image georegistration; S2. Image preprocessing: The acquired visible light images are preprocessed, and image georegistration is performed based on the UAV attitude parameters to eliminate image offset caused by flight attitude fluctuations; The dataset was expanded using an improved Mosaic data augmentation method, while brightness equalization, noise removal and image cropping were performed. The output size was uniformly set to 640×640 pixels to improve the robustness of the model. S3. Tobacco Field Region Segmentation: An improved semantic segmentation network is used to extract the tobacco field region from the preprocessed image. U-Net is used as the base network, and the SCConv module is embedded in the encoder to reduce feature redundancy in the background region and enhance the feature representation of the tobacco field region. At the same time, the CBAM convolutional block attention mechanism is introduced in the decoder to strengthen the extraction of tobacco field boundary features, accurately distinguish between tobacco field and non-tobacco field regions, and output a binarized tobacco field segmentation mask. S4. Tobacco Field Plant Target Detection and Plant Count: Based on the segmented tobacco field area, the improved YOLO-SCAT algorithm is used to detect small targets in the tobacco field plants. YOLOv7 is used as the baseline network, and the last ELAN module of the backbone feature extraction network is replaced with the ELAN-SC module. Furthermore, the CBAM attention mechanism is introduced after the ELAN-SC module. The detection head outputs the predicted bounding box, confidence level, and category information of the tobacco field plants. A confidence level threshold (≥0.5) is set to filter out false detection targets. The detected tobacco field plants are counted to obtain the total number of tobacco field plants. S5. Tobacco field area calculation: The area is calculated by combining the tobacco field segmentation mask and the geographical parameters collected by the UAV. The image pixel resolution is calculated based on the UAV flight altitude and image sensor parameters. Count the total number of foreground pixels in the tobacco field segmentation mask, and calculate the actual area of ​​the tobacco field based on the pixel resolution; S6. Post-processing and output of results: Post-process the results of plant count and area calculation, remove abnormal detection targets in the edge areas, and output plant count and area data and visualization reports for each region in combination with the distribution of tobacco fields.

2. The method for determining the number of tobacco plants (area) in a field based on visible light images from a UAV according to claim 1, characterized in that, Step S1 specifically includes: S1.1 Parameter calculation and equipment calibration: The UAV attitude data heading angle, roll angle and pitch angle are processed by the image geographic coordinate transformation formula to calibrate the shooting angle; S1.2 Image resolution calculation: The image ground sampling distance is determined by the UAV flight altitude H.

3. The method for determining the number of tobacco plants (area) in a field based on visible light images from a UAV according to claim 1, characterized in that, The specific steps of improving the Mosaic data augmentation method in step S2 are as follows: (1) Randomly select N tobacco field images, scale and flip them respectively, with a scaling ratio of 0.5-1.5 times and a horizontal / vertical flip probability of 50%; (2) Normalize the color space of each image to eliminate the influence of lighting differences, and stitch the N normalized images into one 640×640 pixel image. The stitching area adopts linear fusion transition and is stitched in groups of N=4.

4. The method for determining the number of tobacco plants (area) in a field based on visible light images from a UAV according to claim 1, characterized in that, In step S3, the process of embedding the SCConv module at the encoding end based on U-Net is divided into two parts: the Spatial Reconstruction Unit (SRU) is used to separate farmland plants from background features, and the Channel Reconstruction Unit (CRU) is used to reduce the redundancy of similar leaf texture channels.

5. The method for determining the number of tobacco plants (area) in a field based on visible light images from a UAV according to claim 1, characterized in that, In step S4, the CBAM attention mechanism is introduced after the ELAN-SC module, which is a channel attention mechanism and a spatial attention mechanism, to highlight the key channels of the plant and the spatial position of the plant.

6. The method for determining the number of tobacco plants (area) in a field based on visible light images from a UAV according to claim 1, characterized in that, In step S5, the actual area of ​​the tobacco field is related to the number of foreground pixels and the pixel resolution. Furthermore, when the drone has a roll angle or pitch angle, the image will produce perspective distortion, which requires the area to be corrected to obtain the actual area of ​​the tobacco field.

7. The method for determining the number (area) of tobacco plants in a field based on visible light images from a UAV according to claim 1, characterized in that, In step S6, the specific process of removing abnormal detection targets in the edge region is as follows: an "isolated detection box" is set up at the boundary of the tobacco field, and spatial clustering algorithm is used to remove them.

8. The method for determining the number of tobacco plants (area) in a field based on visible light images from a UAV according to claim 7, characterized in that, The specific process of the spatial clustering algorithm for elimination is as follows: (1) Extract the center coordinates of all detection boxes and determine the pixel coordinate range of the paddy field area based on the paddy field segmentation mask; (2) Calculate the minimum distance from the center of each detection box to the paddy field area; (3) Set a distance threshold. If the minimum distance is greater than the distance threshold, the detection box is eliminated.

9. A system for implementing the method for determining the number (area) of tobacco plants in a field based on visible light images from a UAV as described in any one of claims 1-8, characterized in that, The system includes: an image acquisition module, a data preprocessing module, a tobacco field segmentation module, a target detection and counting module, an area calculation module, and a result output module. The image acquisition module consists of a UAV and a visible light image sensor for image acquisition. The data preprocessing module receives image data from the image acquisition module and performs preprocessing. The tobacco field segmentation module accurately segments the tobacco field area from the background area using the preprocessed image data and outputs a tobacco field segmentation mask. The target detection and counting module performs small target detection on tobacco plants in the segmented tobacco field area and outputs the total number of tobacco plants. The area calculation module calculates the actual area of ​​the tobacco field based on the image pixel resolution, the number of foreground pixels in the tobacco field segmentation mask, and the geographical parameters of the UAV. The result output module performs post-processing on the plant count and area data to generate a visual report.

10. A system for determining the number (area) of tobacco plants in a field based on visible light images from a UAV, as described in claim 9, is characterized in that... The tobacco field segmentation module incorporates an improved U-Net semantic segmentation network and utilizes the SCConv module and CBAM attention mechanism. The target detection and counting module deploys an improved YOLO-SCAT algorithm model.

Images