Measuring Agricultural Area Using YOLO Object Detection and ArUco Markers

In this section, a discussion of the research results will be presented at the beginning of the research until conclusions emerge. In accordance with the research stages in chapter three, the initial stages are in the form of data acquisition using drone equipment that flies over rice paddy fields. Data acquisition takes the form of taking images of rice fields at varying distances, namely 2 meters, 5 meters, 10 meters and 20 meters. The purpose of taking images with various distances is to find the best distance in terms of image detection of infected paddy fields and healthy fields. Table 4 shows the time of data acquisition along with the data acquisition distance from the drone camera to the object. The acquisition distance between the drone and the object in Figure 7 shows the image acquisition results. The image acquisition process is carried out perpendicular to 90° it easier to calculate the area based on the detection results of the bounding box.

Conduct training with the aim of training and finding the best model in determining patterns and characteristics of each dataset class or better known as image feature extraction. The existence of a training process is a basis for determining the best decision in predicting each case of healthy and infected images. The training dataset at the start of training uses Yolo v4 pre-trained weights from the darknet until we get the best weights according to the dataset that has been prepared. A technique like this is an implementation of the transfer learning technique, which means that the parameters used have been trained by another party so there is no need to train from scratch. Meanwhile, dataset training is carried out on the Google Collaboratory platform due to its free nature, but for large amounts of data requiring high computing, you need to upgrade to Google Colab Pro or Google Colab Pro+. In Table 4 are the parameters used during dataset training.

Figure 8 depicts the outcomes of a training model encompassing two classes (infected and healthy). The model achieves the best mean Average Precision (mAP) of 77.3% on a 20-meter acquisition image, while the lowest mAP value, 46.8%, is observed at a 2-meter distance. Notably, the largest loss value is recorded in the 2-meter acquisition image, whereas the smallest loss occurs in the 5-meter acquisition image. This experimental data indicates that image acquisition from a 2-meter distance yields the least favorable result among the varied acquisition distances in the dataset.

The performance of the training model across diverse image dataset variations underscores the impact of various factors on image acquisition. These factors include acquisition distance, light intensity during image acquisition, camera specifications, acquisition time, weather conditions, and labeling. The comprehensive results of the training model across different image acquisition distances are detailed in Table 5. Additionally, the swiftest detection time is achieved in 1 second for the labeling of the two specified classes.

Object detection results provide a percentage level indicating the confidence of the detection. Figure 9 illustrates these values for both healthy and infected labels. The detection value influences the prediction outcome, as the training process relies on the feature extraction value for each object. Some detection instances yield a 100% value, signifying a true accuracy rate, while others may show values like 0.97. Variances in detection values among objects highlight distinctions in their unique characteristics. Figure 10 presents results from detection at a 10-meter acquisition distance, indicating that one land is identified as infected, while two lands are identified as healthy.

The next step in this research is to find the pixel area of the reference image, namely the ArUco Marker, at varying shooting distances. The measurement of the image pixel area is carried out to be used as a reference in units of length (cm) per 1 square meter area. So this time we will look for the area of pixels in 1 square meter. Figure 5 shows a reference image that will calculate the pixel area per 1 square meter. Reference images will be acquired using drone cameras at varying distances. The reference image acquisition distance starts from 1 meter to 20 meters with a range of 1 meter. This is because it makes it easier to find mathematical equations as a model for determining the area of an image area based on images obtained by drones.

Figure 5 is a reference image in the form of an ArUco Marker with a real size of 100cm x 100cm. From the image, the length and width of the pixels are searched starting from one corner point to another corner point. Using Euclidean equations (equation 1) and the Matlab tool, you can find the coordinates of each corner, namely {(x_max,y_min), (x_max,y_max), (x_min,y_min), (x_min,y_max)} with coordinate values {(234,106), (1248, 121), (229,1124), (1241,1112)}. The results of the coordinate values are illustrated according to Figure 1 so that the values of the length, width and area of the reference image area will be known.

$\begin{gathered}d=\sqrt{\left(x_{\max }-x_{\max }\right)^2+\left(y_{\max }-y_{\min }\right)^2} \\ \text { Width }=1.014 \text { pixel }\end{gathered}$               (1)

Meanwhile, for calculating length distances in coordinates (234,106), (229,1124)

Length $=\sqrt{\left(x_{\max }-x_{\min }\right)^2+\left(y_{\min }-y_{\min }\right)^2}$

Length $=1.018$ pixel

Table 4. Parameter model Yolo V4

7.png

Figure 7. Results of drone image acquisition

81.png

82.png

83.png

1.png

Figure 8. Result of model training: (a) 2 meter (b) 5 meter (c) 10 meter (d) 20 meter

The pixel area results in the reference image show the image pixel area per 1 square meter at a certain acquisition distance. Thus, when calculating the area of infected and healthy plants, we only compare the image of the rice area with the reference image but must be adjusted to the same acquisition distance. Table 6 shows the results of measuring the pixel area of the reference image acquired from various distances.

After looking for size calculations on reference images at various acquisition distances, we now start calculating the area of infected and healthy rice plants based on the detection results using the Yolo v4 algorithm. Using the detection data in Figure 10, we will look for the area of infected and healthy rice plants by referring to the ArUco Marker reference image. The image obtained from the acquisition in Figure 10 was acquired from a distance of 11 meters so that in calculating the area area it will be referenced with the reference image acquired from a distance of 11 meters. To make calculations easier, the image of the rice plant was converted from a color image to a gray image according to Figure 11. After that, from the gray image, the pixel area was searched by first calculating the length and width of the pixels in the image where the infection was detected using the Euclidean equation. Figure 10 shows the results of the Yolo detection results, which are only in the infected area.

Using the Euclidean equation, the length and width of the image pixels for the infection detected in Figure 10 are calculated. The coordinate point values of the detected image are in the coordinate point format {(x_max,y_min), (x_max,y_max), (x_min,y_min), (x_min,y_max)} the value is known {(1141,15), (2572,15), (1141,2134), (2572,2134)}. To clarify the location of the coordinate points at each corner of the detected image, it can be seen in Figure 10.

Table 7. Calculation results of reference and detected image

Table 8. Calculation results of the pixel area of the reference image and the detected image

Table 9. Comparison of the results of calculating the area of rice plantations

Using the illustration in Figure 10, the pixel length and width values can be calculated with equation 1 as follows:

Length $=2.119$ pixel

Meanwhile, the pixel width value using the same method can be calculated with equation 1 as follows:

Width $=1.431$ pixel

Thus, for the image pixel area, the detection results are:

Square area $=3.032 .289$ pixel

After the length, width and pixel area of the detected image are known, the next step is to reference the reference image at the same acquisition distance, namely 11 meters. Reference pixel areas are presented in Table 6 to determine the real area of the detected image. A summary of the calculation of the length, width and area of the reference image and the detected image is presented in Table 7. To calculate the real area of the detected image, you need to compare the area of the detected image with the area of the reference image because the reference image is the real area per 1 meter².

After knowing the results of the area of the detected image in pixels, it will then be converted into an area in meters. The calculation of the real area of the detected image is as follows:

$\begin{gathered}\text { Length }=\frac{\text { Length detected image }(\text { in pixel })}{\text { Length reference image }\left(\text { in } \frac{\text { pixel }}{m}\right)} \\ \text { Length }=\frac{2119}{264}\end{gathered}$

Length $=8,03$ meter

$\begin{gathered}\text { Width }=\frac{\text { Width detected image }(\text { in pixel })}{\text { Width reference image }\left(\text { in } \frac{\text { Pixel }}{m}\right)} \\ \text { Width }=\frac{1431}{264} \\ \text { Width }=5,42 \text { meter }\end{gathered}$

Square area $=$ Length $x$ Width

Square area $=8,03 \times 5,42$

Square area $=43,51$ meter $^2$

Thus the results of the calculation of the area of the image detected an infection of 43.51 meters².

As an evaluation material for the results of calculating the area of plants resulting from Yolo detection images, calculations will be carried out using different resulting images. Figure 12 shows the image of the detection results which will be used as evaluation material by comparing the calculation results with the actual area of rice plants. The image used in this evaluation measures 11 meters long and 5 meters wide and was acquired from a distance of 10 meters.

Using the Euclidean equation, the length and width of the image pixels detected as healthy in Figure 10 are calculated. The coordinate point values of the detected image using the coordinate point format {(x_max,y_min), (x_max,y_max), (x_min,y_min), (x_min,y_max)} are known the values are {(19,25),(3353,25),(19,1463),(3353,1463)}. To clarify the location of the coordinate points at each corner of the detected image, it can be seen in Figure 13.

12.png

Figure 12. Image of the results of the Yolo detection as an evaluation

13.png

Figure 13. The corner coordinates of the detected image

Using the illustration in Figure 13, the pixel length and width values can be calculated with Eq. (1) as follows

Length $=3.334$ piksel

Meanwhile, the pixel width value using the same method can be calculated with equation 1 as follows:

Width $=1.438$ pixel

Thus, for the image pixel area, the detection results are :

Square area=Length $x$ Width

Square area $=4.794 .292$ pixel

After the length, width and area of the detected image pixels are known, the next step is referenced with a reference image at the same acquisition distance of 10 meters. The pixel area reference is presented in Table 6 to find out the real area of the detected image. A summary of the calculation of the length, width and area of the reference image and detected image is presented in Table 8. To calculate the real area of the detected image is to compare the area of the detected image with the area of the reference image because the reference image is the real area per 1 meter².

The calculation of the real area of the detected image is as follows:

$\begin{gathered}\text { Length }=\frac{\text { Length detected image }(\text { in pixel })}{\text { Length reference image }\left(\text { in } \frac{\text { pixel }}{m}\right)} \\ \text { Length }=\frac{3.334}{291} \\ \text { Length }=11,45 \text { meter }\end{gathered}$

$\begin{gathered}\text { Width }=\frac{\text { Width detected image }(\text { in pixel })}{\text { Width reference image }\left(\text { in } \frac{\text { pixel }}{m}\right)} \\ \text { Width }=\frac{1.438}{291} \\ \text { Width }=4,94 \text { meter }\end{gathered}$

Square area $=$ Length $x$ Width

Square area $=56,57$ meter $^2$

Thus, the real area of the image detected as healthy in Figure 13 using manual calculations has a length of 11.45 meters and a width of 4.94 meters and the real area of the image detected is 56.57 meters2. The results of manual calculations using the ArUco Marker reference image on the Yolo v4 algorithm and measuring the real area of the plant area produce almost the same values. Table 9 shows a recapitulation of the results of area calculations using the ArUco Marker and the actual area of rice plants detected.

Table 9 shows the results of calculations using the ArUco Marker reference image with the actual area of infected rice land producing almost the same values. Calculations using the ArUco Marker reference image produce a value of 56.67 m² and a real area of 55 m². If you look at the calculation results, the accuracy of the calculation using the ArUco Marker reference image is 97.05%. Thus, the use of the ArUco Marker reference image can be used as material for measuring area size based on the image obtained by the drone camera.