Rapid and efficient counting method suitable for early seedling emergence stage of cotton
Patent Information
- Application Number
- CN202311454183.X
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2023-11-03
- Publication Date
- 2025-05-06
AI Technical Summary
The prior art has problems in the counting of early cotton seedlings with large prior data demands and high environmental conditions, which leads to unstable counting accuracy, especially in complex weather conditions, which is difficult to effectively monitor.
The waveform method is combined with the drone RGB image, and through ExG index binarization, linear detection and masking treatment, cotton seedling rows are extracted and waveform extraction and smoothing are performed, and the wave peaks are located and counted to achieve fast and efficient counting of cotton seedlings.
This method significantly reduces the requirements for image resolution and environmental conditions under different environmental conditions, improves the accuracy and stability of cotton seedling counting, and can achieve efficient and economical monitoring in the absence of empirical data.
Smart Images
Figure CN119942319A_ABST
Abstract
Description
Technical Field
[0001] The invention belongs to the field of non-destructive detection of crop seedling status based on RGB sensors at the scale of unmanned aerial vehicles, and particularly relates to a fast and efficient counting method for early emergence of cotton seedlings suitable for equally spaced row-sown crops. Background Art
[0002] Counting large areas of crop seedlings in the field is time-consuming and laborious. It is considered to be highly subjective and cannot provide a macro and accurate description of field seedling emergence. With the rise of drone technology, the use of drones to monitor seedling numbers has been widely used in crops such as corn, rapeseed, wheat, and cotton. In the early stages of seedling monitoring, RGB sensors, multispectral sensors, hyperspectral sensors, and near-infrared sensors are often used for counting. The counting methods are mainly divided into: (1) methods based on canopy morphological features extracted from spectral features or combining geometric features (area, perimeter, etc.) to establish a regression relationship with the actual number; (2) deep learning image recognition counting methods based on a large amount of prior training data and labels.
[0003] Among them, the counting method based on regression method completes the estimation of seedling number by establishing a regression model between vegetation information obtained from drone images and the actual number of seedlings. In the regression model, vegetation information is obtained by separating vegetation from the background in the original image using threshold method, classification method, or vegetation index (such as ExG, NDVI (Normalized Difference Vegetation Index), EVI (Enhanced Vegetation Index) or GNDVI (Green Normalized Difference Vegetation Index)). Accurate extraction of vegetation features is the basis for subsequent geometric feature extraction. In order to extract geometric features, OSTU, supervised classification, K-means classification, etc. are often used to binarize the image to obtain vegetation coverage area, centroid, aspect ratio, number of polygons, perimeter, and other geometric information. Finally, the early canopy geometric or morphological characteristic parameters obtained by the comprehensive vegetation index and image processing method are correlated with the field survey seedling number data on the corresponding date to establish an estimation model containing geometric and morphological characteristic parameters and the measured number of seedlings. The segmentation results of seedlings and background are greatly affected by light, and image data must be obtained under clear and windless conditions; the above regression method has high requirements on image resolution and is only used for fixed emergence dates. Different monitoring models need to be established for different emergence dates, and its promotion and applicability is poor; differences in weather, light and image resolution when acquiring images will increase the uncertainty of extraction accuracy; in reality, rainy weather cannot be avoided in the early stage of emergence, which will inevitably affect the monitoring effect of geometric morphology, thereby affecting the counting effect during the emergence process.
[0004] In addition to the regression-based counting method, with the development of computer technology, deep learning technology has also been widely used in seedling counting. This method learns the features of the labeled target in the image to more precisely identify the target and count it. The simplified VGG16 and U-Net CNN were used for rice and sorghum ear detection and counting, respectively. The use of the YOLOv3 object detection model for cotton seedling positioning and counting can eliminate the need for geometric and statistical information. In addition, in the study of the impact of different environments on target detection, the improved ReCNN was used to train more than 20,000 corn seedling images collected in complex environments, and it was found that the accuracy of target detection under sunny days was lower than that under cloudy days. The deep learning method reduces the environmental conditions when the image is acquired, but when the resolution decreases, the seedlings are too small and difficult to distinguish from the background, and the image resolution is still required to be high. In order to solve the problem of high resolution requirements, the MaxArea mask score and RCNN algorithm are used to process the image by row, which not only reduces the requirements for crop clarity in the image, but also has robustness in the application of data collected in different years. The accuracy of the estimated seedling emergence rate is as high as 95.8%. Although deep learning does not require the collection of a large amount of field prior data, it does require the labeling of data from different data sets, which places high demands on computer performance.
[0005] The canopy size of a single cotton seedling is large and independent, and the appearance of the canopy changes greatly during growth. Therefore, when counting seedlings, this method needs to be insensitive to morphological changes to cope with the morphological changes of cotton seedlings at different detection times; insensitive to changes in light to cope with the changeable weather in the early stage of seedling emergence. We also tried to monitor the early emergence of seedlings without the data labels of the year to ensure that early monitoring after emergence can ensure early seedling replenishment. Summary of the invention
[0006] The purpose of the present invention is to propose a low-cost, high-efficiency counting method suitable for early cotton seedling monitoring. The method can solve the problems of large demand for prior data and high requirements for environmental conditions for early seedling counting, and ensure that satisfactory counting accuracy can be achieved under low-cost conditions.
[0007] To achieve the above object, the technical solution adopted by the present invention is: a fast and efficient counting method suitable for the early stage of cotton seedling emergence, the steps of which are as follows:
[0008] Step 1: Data collection: Obtain UAV RGB images of cotton canopy at different days after sowing in the early stage of cotton emergence;
[0009] Step 2: Output orthophoto: Import the collected RGB image into the image stitching software, and output the orthophoto after completing the image stitching;
[0010] Step 3: Extracting the seedling rows: After calculating the ExG (excess green) index, binarize the image, use the line detection method to obtain the crop rows, expand the crop rows to generate the mask boundary, and extract the cotton seedling rows after masking;
[0011] Step 4: Use the waveform method to count cotton seedlings: The DN values of the pixels perpendicular to the crop row in the ExG index image are accumulated to obtain a waveform curve with multiple peaks and troughs. The peaks that meet the requirements are located and counted, and the seedlings in the row are located and counted.
[0012] Step 5: Verify the accuracy of cotton seedling counting using the waveform method: Use field measured data or manual counting from images as the standard and use RMSE and R 2 Two indicators were used to evaluate the WM counting effect.
[0013] Furthermore, in step 1, data is collected from drone data of different dates in the early stage of cotton emergence, wherein the drone collection steps are as follows:
[0014] a. Plan the flight area and route: Use DJI GS PRO to plan the vector boundary of the target area;
[0015] b. Set shooting and drone flight parameters: Use DJI GO 4 to set flight parameters, including a flight altitude of no less than 10m, and sideways overlap and row overlap rates both must be greater than 75%;
[0016] c. Mavic 2 was used to collect images of the cotton canopy at different days after sowing in the early emergence stage.
[0017] Furthermore, in step 2, the image is resampled and the brightness is adjusted, including:
[0018] a. Image stitching: Import the canopy dataset into Agisoft PhotoScan, and output the orthophoto in tiff format after image alignment, high-density point cloud construction, mesh structure, texture structure, tiled building model, and DEM construction;
[0019] b. Cropping and resampling: Define the projection coordinates for the tiff format drone image, crop the image range in the ENVI software, and use the nearest neighbor interpolation method to resample to a certain resolution with a certain step size;
[0020] c. Change image brightness: adjust the I in the HSI color model of the image, expand and reduce it with a certain step size;
[0021] d. Use visual interpretation methods in orthophotos to manually mark and count cotton seedlings in the images. To reduce the subjective differences in manual counting, the data marked by at least three people are combined as the final result.
[0022] Furthermore, in step b, the image is resampled to a resolution of 2.5 mm with a step size of 0.1; in step c, the image is enlarged and reduced with a step size of 0.1, and the multiples of the enlargement and reduction are 0.8, 0.9, 1.1, and 1.2 respectively.
[0023] Furthermore, in step 3, the crop rows are obtained using a straight line detection method, and then the influence of weeds is removed by a masking method while reducing the amount of calculated data, leaving only the cotton seedling rows of interest, specifically including:
[0024] a. Extract green information: After obtaining the orthophoto, use enhanced green information to enhance the green information in the image:
[0025] ExG=2×GRB
[0026] Among them, R, G, and B represent the red, green, and blue bands respectively;
[0027] b. Line detection: To ensure that a straight line is accurately detected for each seedling row, the binary image is expanded at an angle of α to connect the cotton seedlings in the row direction. Then, the Hough transform line detection is performed on the binary image after the seedlings are connected. In this way, a straight line passing through all the seedlings can be detected for each seedling row.
[0028] c. Masking planting rows: After Hough transform straight line detection, the coordinates of the two endpoints of the center line segment of the planting row, point1 and point2, can be determined, and the equation of the line segment can be obtained: y=a*x+b; set the planting row width to w, take the seedling row as the center line, and expand b1 and b2 to the left and right respectively to determine the area of the planting row, calculate the position information in the image, and for any point that satisfies y-ax+b1>0 and y-ax+b2<0 at the same time, the pixel at the point is assigned a value of 1, and the values of the remaining areas are zero; obtain the intersection of the left and right areas of the straight line to obtain the planting row area where the straight line is located, calculate the seedling row detected in the input image and then find the union, which can quickly and efficiently complete the planting row masking of the image.
[0029] Furthermore, the relationship between the left and right extensions b1 and b2 and the planting row width w is set to |b1|=|b2|=w / 2.
[0030] Furthermore, the step 4 specifically includes:
[0031] a. Waveform extraction: The crop row ExG image is obtained by masking the ExG image with the crop row. In this image, the DN values of the pixels perpendicular to the crop row are accumulated in units of rows to obtain the waveform curve of the row.
[0032] b. Waveform smoothing: Use Gaussian smoothing method to smooth the extracted waveform curve;
[0033] c. Peak extraction: Extract the peaks from the smoothed waveform curve.
[0034] Furthermore, in step c, two characteristic parameters need to be set to extract the peaks: the lowest peak value and the distance between the two peaks;
[0035] First, set the minimum peak value to ignore background noise. The minimum value must be greater than the accumulation of pure soil pixel DN values. Data below the base value line will not be included in the peak selection.
[0036] The second step is to set the distance between the two peaks, which is half the planting distance. This is used to filter out peaks that are too close and reduce miscounts caused by the shape of the seedlings.
[0037] Furthermore, the peak value is extracted using the local maximum method. The pixel position corresponding to the peak is the seedling position, and the cotton seedling counting is completed by counting the peaks.
[0038] Furthermore, in step 5, RMSE is used to measure the deviation between the observed value and the true value, using R 2 To measure the stability of the counting method:
[0039]
[0040]
[0041] Where: y i represents the ground truth value of the i-th row; represents the estimated quantity of the i-th row; represents the average number of seedlings in each seedling row in the figure; m represents the number of rows.
[0042] Compared with the prior art, the present invention has the following beneficial effects:
[0043] The waveform method is based on the characteristics of the UAV, the study area and the research crops. It has low requirements on UAV performance, flight environment and image resolution, which is beneficial to the promotion of UAV remote sensing seedling monitoring without empirical data.
[0044] The advantage of the counting method of the present invention is that when there is a lack of empirical data, it is a more economical and convenient counting method that is less restricted by environmental factors. It is an efficient, precise and automatic counting method for cotton seedlings to cope with complex weather. It provides methods and ideas for the precise counting of equally spaced row-sown crops based on drone images. BRIEF DESCRIPTION OF THE DRAWINGS
[0045] Figure 1Schematic diagram of cotton seedling counting process based on waveform method.
[0046] Figure 2 Extract schematic diagrams for planting rows.
[0047] Figure 3 The original waveform (a), smoothed waveform (b), and peak extraction (c).
[0048] Figure 4 The results of seedling counting on different dates are shown below. The seedling size in the image obtained on May 7 was the smallest. When counting 10m long and 12 rows, the counting effect RMSE = 1.8708, R 2 =0.9541; The image was acquired on May 12, with the same plant spacing of 10 cm, but it was a cloudy day and the seedling canopy size was larger, so the counting accuracy remained stable, RMSE = 2.121, R 2 =0.9602.
[0049] Figure 5 The counting effect at different resolutions. The counting was monitored 7 days after sowing. There was no obvious change trend in the two indicators within the resolution range of 0.33-1.0 cm. The RMSE was less than 2.0 plants / row, ranging from 1.55-1.93, and R2 was stable at 0.96-0.95. After the resolution exceeded 1.0 cm, the RMSE fluctuated and increased, ranging from 2.0 to 5.2 plants / row in the range of 1.0-2.4 cm, and the R2 fluctuated and decreased, ranging from 0.79 to 0.93. When the resolution was 2.5 cm, the accuracy dropped rapidly, with R2 of 0.74 and RMSE of 5.22 plants / row. Monitoring was conducted 12 days after sowing. Between 0.33 and 1.4 cm, RMSE was stably distributed in the range of 1.7-2.3, and R2 was stably distributed in the range of 0.94-0.97. At 1.3 cm, the accuracy was the best, with RMSE of 1.08 plants / row and R2 of 0.99. At 1.5-2.4 cm, RMSE fluctuated upward in the range of 2.4-4.4, and R2 fluctuated downward slightly in the range of 0.90-0.95. Similarly, at a resolution of 2.5, the accuracy dropped rapidly, with RMSE of 6.3 plants / row and R2 of 0.7.
[0050] Figure 6 Counting results at different resolutions 7 days after seeding. 7 days after sowing, the counting results were not much different when the resolution was below 1.0 cm. When the resolution was greater than 1 cm, the counting accuracy changed greatly. When the resolution was 2.0-2.5 cm, it can be seen from the figure that the resolution at this time could no longer accurately capture the seedling information. The maximum resolution of monitoring 7 days after sowing was kept within 2.0 cm, which can ensure R 2 greater than 0.8, and RMSE less than 4 plants / row.
[0051] Figure 7Counting results at different resolutions 12 days after emergence. For monitoring 12 days after emergence, 1.4 cm resolution is the most suitable under the condition of ensuring accuracy and efficiency. The maximum RMSE difference is 2 plants / row when the resolution is in the range of 1.4-2.4 cm. Underestimation occurs at 2.3 cm resolution 12 days after sowing. The resolution of overall underestimation 12 days after sowing is coarser than that of overall underestimation 7 days after sowing.
[0052] Figure 8 The results of counting at different brightnesses 7 days after seedling emergence. The accuracy of the counting results decreased when the brightness was reduced by 0.8 and 0.9 times, and the R2 and RMSE were 0.92, 0.88, 2.33 plants / row, and 2.90 plants / row, respectively. When the brightness increased to 1.1 and 1.2 times, the underestimation of the number of seedlings was weakened, the scatter points were closer to the 1:1 line, and the calculation accuracy was improved, with R2 and RMSE being 0.95, 0.96, 1.89 plants / row, and 1.66 plants / row, respectively.
[0053] Fig. 9 The results of counting at different brightness levels 12 days after seedling emergence. When the brightness increases to 1.2, it is the same as the brightness of the image obtained on a sunny day. When the brightness is reduced by 0.8, 0.9, and 1.1 times, it is lower than the brightness on a sunny day. Both R2 and RMSE remain stable, with R2 between 0.95-0.97 and RMSE between 2.02-2.22 plants / row.
[0054] Fig.10 Counting results for different brightness. DETAILED DESCRIPTION
[0055] The present invention is described in detail below with reference to the accompanying drawings and specific embodiments.
[0056] Figure 1 As shown, a fast and efficient counting method suitable for the early stage of cotton emergence is provided, and the specific steps are as follows:
[0057] Step 1: Data collection: Obtain UAV RGB images of cotton canopy at different days after sowing in the early stage of cotton emergence.
[0058] include:
[0059] a. Plan the flight area and route: Use DJI GS PRO to plan the vector boundary of the target area;
[0060] b. Set the shooting and drone flight parameters: Use DJI GO 4 to set the flight parameters, including the flight altitude of 10m, and the lateral overlap rate and the row overlap rate must be greater than 75%;
[0061] c. Mavic 2 was used to collect images of the cotton canopy at different days after sowing in the early emergence stage.
[0062] Step 2: Output orthophoto: Collect RGB images and import them into image stitching software. After image stitching, output orthophoto. In order to study the application of this method at different resolutions and brightness, resample and adjust the brightness of the images. Including:
[0063] a. Image stitching: Import the canopy dataset into Agisoft PhotoScan, and output the orthophoto in tiff format after image alignment, high-density point cloud construction, mesh structure, texture structure, tiled building model, and DEM construction;
[0064] b. Cropping and resampling: Define the projection coordinates for the tiff format drone image, crop the image range in the ENVI software, use the nearest neighbor interpolation method with a step size of 0.1, and resample to a resolution of 2.5 mm;
[0065] c. Change image brightness: adjust the I in the HSI color model of the image, expand and reduce it with a step size of 0.1. The multiples of expansion and reduction are 0.8, 0.9, 1.1, and 1.2 respectively;
[0066] d. Use visual interpretation methods in orthophotos to manually mark and count cotton seedlings in the image. To reduce the subjective differences in manual counting, the data marked by three people are combined as the final result.
[0067] Step 3: Extracting the rows of seedlings: After calculating the ExG index, binarize the image, use the line detection method to obtain the crop rows, and then use the masking method to remove the influence of weeds and reduce the amount of calculated data, leaving only the rows of cotton seedlings in the area of interest. Specifically include:
[0068] a. Extract green information: After obtaining the orthophoto, use green plants and soil to distinguish. Here, the common enhanced green information (ExG index) is used to enhance the green information in the image:
[0069] ExG=2×GRB
[0070] R, G, and B represent the red, green, and blue bands respectively;
[0071] b. Line detection: To ensure that a straight line is accurately detected for each seedling row, the binary processed image is expanded at an angle α to connect the cotton seedlings in the row direction ( Figure 3 b), then performing straight line detection on the binary image after the seedlings are connected, so that a straight line passing through all the seedlings can be detected in each seedling row;
[0072] c. Masking planting rows: After Hough transform straight line detection, the coordinates of the two endpoints of the center line segment of the planting row, point1 and point2, can be determined, and the equation of the line segment can be obtained: y = a*x + b. Set the planting row width to w (the value is determined by the pixel size and the plant spacing, and this article sets it to 1.5 times the plant spacing), take the seedling row as the center line, and expand b1 and b2 to the left and right to determine the area of the planting row (|b1| = |b2| = w / 2), calculate the position information in the image, and for any point that satisfies y-ax+b1>0 and y-ax+b2<0 at the same time, the pixel at the point is assigned a value of 1, and the values of the remaining areas are zero. Obtaining the intersection of the left and right areas of the straight line can obtain the planting row area where the straight line is located. After calculating the detected seedling rows in the input image and finding the union, the planting row mask of the image can be completed quickly and efficiently.
[0073] Step 4: Use the waveform method to count cotton seedlings: The DN values of the pixels perpendicular to the crop row in the ExG image are accumulated to obtain a waveform curve with multiple peaks and troughs. The peaks that meet the requirements are located and counted, and the seedlings in the row are located and counted. This includes:
[0074] a. Waveform extraction: The crop row ExG image is obtained by masking the ExG image with the crop row. In this image, the DN values of the pixels perpendicular to the crop row are accumulated in units of rows to obtain the waveform curve of the row.
[0075] b. Waveform smoothing: The waveform curve extracted using the Gaussian smoothing method is smoothed;
[0076] c. Peak extraction: extract the peaks from the smoothed waveform curve. Peak extraction requires setting two characteristic parameters: the lowest peak value and the distance between the two peaks. First, set the lowest peak value. Its purpose is to ignore background noise. The minimum value must be greater than the accumulated DN value of the pure soil pixel (base value). Data below the base value line does not participate in peak selection; secondly, set the distance between the two peaks. The distance between the two peaks is half of the planting distance. The purpose is to filter out peaks that are too close and reduce miscounts caused by seedling shape. Peak extraction uses the local maximum method. The pixel position corresponding to the peak is the seedling position. Counting the peaks completes the cotton seedling counting.
[0077] Step 5: Verify the accuracy of cotton seedling counting using the waveform method: Use field measured data or manual counting from images as the standard and use RMSE and R 2 Two indicators are used to evaluate the WM counting effect. RMSE is used to measure the deviation between the observed value and the true value. 2 The counting method stability was measured.
[0078]
[0079]
[0080] where y i represents the ground truth value of the i-th row; represents the estimated quantity of the i-th row; represents the average number of seedlings in each seedling row in the figure; m represents the number of rows.
[0081] The above method is further described below in conjunction with specific embodiments:
[0082] This example is based on drone data from different plots and different dates and the corresponding measured seedling numbers, as shown in Table 1:
[0083] Table 1 UAV data acquisition
[0084]
[0085] The cotton seedling counting study was conducted based on the drone data collected at the Shihezi University Agricultural Experiment Station in Shihezi City, Xinjiang on May 7 and May 12, 2020. The cotton variety was Lumianyan 24, and two plots of 10m×10m were selected, numbered Ⅰ and Ⅱ. The two plots were used for nitrogen fertilizer application experiments. They were sown at the same time on April 30, with a plant spacing of 10cm. Variable fertilization operations had not yet been carried out in the early stage of emergence. Except for the geographical location, the other conditions of the two plots were the same.
[0086] Both drone images were acquired at noon, with the difference being that May 7 was clear and cloudless, while May 12 was cloudy. The drone equipment used was the Mavic 2 drone produced by Shenzhen DJI Innovations Technology Co., Ltd. (DJI, Shenzhen, Guangdong, China). The original data resolution was 0.33 cm. For the convenience of calculation, it was resampled to 0.4 cm and then resampled to 2.5 cm in steps of 0.1 to perform stability tests of waveform counting at different resolutions. In addition, the brightness of the original image was adjusted, and the multiples of I in the HIS model were adjusted to perform stability tests at different brightness levels.
[0087] The counting results are as follows Figure 4-10 As shown, 7 days after sowing, when the resolution was less than 1.0 cm, the RMSE was less than 2.0 plants / row, and R2 was stable at 0.96-0.95; when the resolution exceeded 1.0 cm, the RMSE and R2 fluctuated greatly; 12 days after sowing, the counting effect was stable under different conditions, and the RMSE was 1.7-4.4 plants / row at different resolutions, and R2 was smoothly distributed in the range of 0.90-0.97. Under different brightness, R2 was between 0.95-0.97, and RMSE was between 2.02-2.22 plants / row.
[0088] In summary, the waveform method is used in this embodiment to monitor the emergence of cotton seedlings, and tests are carried out on images with different emergence dates, resampling and brightness adjustment. It is proved that the counting results are less affected by image resolution and brightness changes, and can be accurately counted while having wide applicability and high efficiency. It is an efficient, precise and automatic counting method for cotton seedlings in complex weather.
[0089] The above shows and describes the basic principles, main features and advantages of the present invention. Those skilled in the art should understand that the above embodiments do not limit the protection scope of the present invention in any form, and all technical solutions obtained by equivalent replacement and the like fall within the protection scope of the present invention. The parts not involved in the present invention are the same as the prior art or can be implemented by the prior art.
Claims
1. A fast and efficient counting method suitable for the early stage of cotton emergence, characterized in that Here are the steps: Step 1: Data collection: Obtain UAV RGB images of cotton canopy at different days after sowing in the early stage of cotton emergence; Step 2: Output orthophoto: Import the collected RGB image into the image stitching software, and output the orthophoto after completing the image stitching; Step 3: Extracting the seedling rows: After calculating the ExG index, binarize the image, use the line detection method to obtain the crop rows, expand the crop rows to generate the mask boundary, and extract the cotton seedling rows after masking; Step 4: Use the waveform method to count cotton seedlings: The DN values of the pixels perpendicular to the crop row in the ExG index image are accumulated to obtain a waveform curve with multiple peaks and troughs. The peaks that meet the requirements are located and counted, and the seedlings in the row are located and counted. Step 5: Verify the accuracy of cotton seedling counting using the waveform method: Use field measured data or manual counting from images as the standard and use RMSE and R 2 Two indicators were used to evaluate the counting effect of waveform method.
2. A fast and efficient counting method suitable for early cotton emergence according to claim 1, characterized in that: In step 1, data is collected from drone data on different dates in the early stage of cotton emergence, and the drone collection steps are as follows: a. Plan the flight area and route: Use DJI GS PRO to plan the vector boundary of the target area; b. Set shooting and drone flight parameters: Use DJI GO 4 to set flight parameters, including a flight altitude of no less than 10m, and sideways overlap and row overlap rates both must be greater than 75%; c. Mavic 2 was used to collect images of the cotton canopy at different days after sowing in the early emergence stage.
3. A fast and efficient counting method suitable for early cotton emergence according to claim 1, characterized in that: In the step 2, the image is resampled and the brightness is adjusted respectively, including: a. Image stitching: Import the canopy dataset into Agisoft PhotoScan, and output the orthophoto in tiff format after image alignment, high-density point cloud construction, mesh structure, texture structure, tiled building model, and DEM construction; b. Cropping and resampling: Define the projection coordinates for the tiff format drone image, crop the image range in the ENVI software, and use the nearest neighbor interpolation method to resample to a certain resolution with a certain step size; c. Change image brightness: adjust the I in the HSI color model of the image, expand and reduce it with a certain step size; d. Visual interpretation methods were used in orthophotos to manually label and count cotton seedlings in the images. To reduce the subjective differences in manual counting, the labeling data of at least three people were combined as the final result.
4. A fast and efficient counting method suitable for early cotton emergence according to claim 3, characterized in that: In step b, the image is resampled to a resolution of 2.5 mm with a step size of 0.1; in step c, the image is enlarged and reduced with a step size of 0.1, and the magnifications of the enlargement and reduction are 0.8, 0.9, 1.1, and 1.2, respectively.
5. A fast and efficient counting method suitable for early cotton emergence according to claim 1, characterized in that: In the step 3, the crop rows are obtained by using a straight line detection method, and then the influence of weeds is removed by a masking method while reducing the amount of calculated data, leaving only the cotton seedling rows of interest, specifically including: a. Extract green information: After obtaining the orthophoto, use enhanced green information to enhance the green information in the image: ExG=2×GRB Among them, R, G, and B represent the red, green, and blue bands respectively; b. Line detection: To ensure that a straight line is accurately detected for each seedling row, the binary image is expanded at an angle of α to connect the cotton seedlings in the row direction. Then, the Hough transform line detection is performed on the binary image after the seedlings are connected. A straight line passing through all the seedlings can be detected for each seedling row. c. Masking planting rows: After Hough transform straight line detection, the coordinates of the two end points of the center line segment of the planting row, point1 and point2, can be determined, and the equation of the line segment can be obtained: y = a*x + b; set the planting row width to w, take the seedling row as the center line, and expand b1 and b2 to the left and right respectively to determine the area of the planting row, calculate the position information in the image, and for any point that satisfies y-ax+b1>0 and y-ax+b2<0 at the same time, the pixel at the point is assigned a value of 1, and the values of the remaining areas are zero; obtain the intersection of the left and right areas of the straight line to obtain the planting row area where the straight line is located, calculate the emergence row detected in the input image and then find the union, which can quickly and efficiently complete the planting row masking of the image.
6. A fast and efficient counting method suitable for early cotton emergence according to claim 5, characterized in that: The relationship between the left and right extensions b1 and b2 and the planting row width w is set to: |b1|=|b2|=w / 2.
7. A fast and efficient counting method suitable for early cotton emergence according to claim 1, characterized in that: The step 4 specifically includes: a. Waveform extraction: The crop row ExG image is obtained by masking the ExG image with the crop row. In this image, the DN values of the pixels perpendicular to the crop row are accumulated in units of rows to obtain the waveform curve of the row. b. Waveform smoothing: Use Gaussian smoothing method to smooth the extracted waveform curve; c. Peak extraction: Extract the peaks from the smoothed waveform curve.
8. A fast and efficient counting method suitable for early cotton emergence according to claim 7, characterized in that: In step c, two characteristic parameters need to be set to extract the peak: the lowest peak value and the distance between the two peaks; First, set the minimum peak value to ignore background noise. The minimum value must be greater than the accumulation of pure soil pixel DN values. Data below the base value line will not be included in the peak selection. The second step is to set the distance between the two peaks, which is half the planting distance. This is used to filter out peaks that are too close and reduce miscounts caused by the shape of the seedlings.
9. A fast and efficient counting method suitable for early cotton emergence according to claim 8, characterized in that: Peak extraction uses the local maximum method. The pixel position corresponding to the peak is the seedling position. Counting the peaks completes the cotton seedling counting.
10. A fast and efficient counting method suitable for early cotton emergence according to claim 1, characterized in that: In step 5, RMSE is used to measure the deviation between the observed value and the true value, and R 2 To measure the stability of the counting method: Where: y i represents the ground truth value of the i-th row; represents the estimated quantity of the i-th row; represents the average number of seedlings in each seedling row in the figure; m represents the number of rows.