Mapping Gaps in Sugarcane Fields in Unmanned Aerial Vehicle Imagery Using YOLOv5 and ImageJ

Abstract

Sugarcane plays a pivotal role in the Brazilian economy as a primary crop. This semi-perennial crop allows for multiple harvests throughout its life cycle. Given its longevity, farmers need to be mindful of avoiding gaps in sugarcane fields, as these interruptions in planting lines negatively impact overall crop productivity over the years. Recognizing and mapping planting failures becomes essential for replanting operations and productivity estimation. Due to the scale of sugarcane cultivation, manual identification and mapping prove impractical. Consequently, solutions utilizing drone imagery and computer vision have been developed to cover extensive areas, showing satisfactory effectiveness in identifying gaps. However, recognizing small gaps poses significant challenges, often rendering them unidentifiable. This study addresses this issue by identifying and mapping gaps of any size while allowing users to determine the gap size. Preliminary tests using YOLOv5 and ImageJ 1.53k demonstrated a high success rate, with a 96.1% accuracy in identifying gaps of 50 cm or larger. These results are favorable, especially when compared to previously published works.

1. Introduction

Sugarcane was the initial crop during the early days of Brazilian colonization, and today, it remains a pivotal component of the country’s economy. Nearly all regions across Brazil engage in sugarcane cultivation, with the state of São Paulo prominently leading the nation in production. Brazil holds the global distinction as the largest sugar producer and the second-largest ethanol producer [1]. Consequently, sugarcane plays a crucial role in fostering economic growth, generating employment opportunities, and contributing substantially to the country’s foreign exchange reserves [2].

Sugarcane stands out as a semi-perennial crop, setting it apart from annual crops. Unlike most crops that follow an annual planting and harvesting cycle, sugarcane boasts a longer life span, typically ranging from three to six years before it is necessary to establish a new sugarcane crop; i.e., sugarcane fields are renewed, on average, every five years [3].

This semi-perennial characteristic arises from its capacity to regenerate and yield stalks after harvest, thus ensuring its sustained presence in the field. However, maintaining the health and productivity of the crop over the years demands effective management practices, including proper fertilization, control of pests and diseases, and precise harvest management. Cultivating sugarcane also entails a substantial initial investment, particularly in tasks such as land preparation, planting, and the initial phases of crop development [4].

Because it lasts for several years, farmers must consider and avoid planting failures in sugarcane fields, because this negatively interferes with the productivity of the crop over these several years of the life of the sugarcane field. The crop’s economic return depends on good plantation maintenance, and it is also affected by the presence of planting failures [5].

Among the factors that cause planting failures are the attack of pests such as nematodes, termites, and sugarcane boll weevil, among others. In addition, failures can be caused by weed infestations, trampling on the stumps, and miscalibrated or worn machinery in harvesting operations [6]. The identification of field gaps allows for the estimation of production breaks, and if gaps exceed 10% [7], gap filling operations are conducted by sugarcane producers [8] to maintain the planned planting density [9]. Gaps are considered empty spaces, measured from center-to-center of the culms at ground level, that are greater than 50 cm [10].

Because sugarcane fields occupy large areas, manual identification of field gaps must be performed by sampling and without creating location maps, in addition to possible bias problems caused by the sampling system. Therefore, there are some works that identify field gaps based on an analysis of the imagery captured from unmanned aerial vehicles (UAVs) [11,12]. UAVs cover large areas in a short time, justifying their use. UAV imagery differs from satellite imagery, having sufficient spatial and temporal resolution to identify various objects of interest, such as gaps in sugarcane crops. UAVs are also affordable for most producers, especially when compared with aircraft flights that are excessively expensive [13,14].

In this study, we utilized the deep learning neural network YOLOv5 [15] due to its lower computational demands compared to R-CNNs and because it also has a high accuracy. The neural network can identify gaps of any size, including those smaller than 50 cm. Subsequently, using ImageJ software, rural producers can determine the size of the field gaps to be displayed on the map in a parameterized manner.

2. Literature Review

2.1. Computer Vision and Object Detection

Computer vision is a multidisciplinary research area within artificial intelligence and machine learning that aims to enable computers to interpret and understand visual information [16]. Object detection, a critical aspect of computer vision, has garnered significant attention [17,18].

Deep neural networks (DNNs) are a type of neural network that is widely employed in computer vision, primarily for identifying objects in images. Convolutional neural networks (CNNs), a specialized form of DNNs, are frequently used to process and analyze visual data [16]. CNNs excel at automatically extracting features from images, which are then used to create models for object classification [19].

2.2. YOLO and Object Detection Techniques

Object detection systems are generally categorized into two groups: two-stage and single-stage detectors. Two-stage detectors, which have a more complex architecture, first identify regions of interest and then pass these regions to a convolutional neural network for further analysis. Single-stage detectors, on the other hand, identify objects in one step using a simpler architecture [18].

Typically, two-stage detection systems, such as R-CNNs, achieve a higher accuracy than single-stage systems. However, single-stage detectors perform faster detections. Notably, the advent of YOLO (you only look once) and its subsequent iterations has significantly enhanced the accuracy of single-stage systems, sometimes surpassing that of two-stage systems [20].

Both YOLO and Faster R-CNN are transformative in precision agriculture for tasks like detecting and monitoring crop pests and diseases, thereby aiding in resource optimization and management. While both systems are effective for various environmental monitoring tasks, YOLO’s flexible architecture makes it particularly adaptable to specific needs [21].

Although R-CNNs generally offer better accuracy compared to YOLO, the latter also provides substantial accuracy and can sometimes outperform R-CNNs. In this study, YOLO was chosen, because R-CNNs require more advanced computational resources, such as state-of-the-art GPUs, making it less accessible for many users [18,20].

2.3. ImageJ for Image Processing

ImageJ is a Java-based platform for image processing, developed by Wayne Rasband at the National Institutes of Health (NIH). Since its launch in 1997, it has been widely utilized in various image-processing applications [22].

While object detection systems like YOLO are designed for identifying and counting objects in images, they lack the capability of detailed image manipulation provided by software like ImageJ. ImageJ offers a range of features, including measuring line lengths, calculating areas, and identifying and counting objects in images. It also supports advanced operations like filtering and performing mathematical operations on pixel values. For color images, these operations are carried out channel by channel, such as the red, green, and blue channels in RGB images [23].

One notable feature of ImageJ is its ability to generate customizable reports that are tailored to specific applications. In this study, it was used to map planting failures in images. The location of an object within an image can be specified in pixels, such as the object’s centroid or the top-left pixel. Additionally, pixel measurements can be converted to other units, such as meters or centimeters, provided that a standard scale is available [23].

2.4. Related Works

In [14], the authors used the linear discriminant analysis (LDA) technique to assess the quality of a sugarcane crop on a farm in Nicaragua, achieving an overall accuracy of 92.9%.

In [11,12], the authors tested the performance of the Inforow software (https://inforow.com.br/en) in detecting planting failures using pixel sizes of 3.5, 6.0, and 8.2 cm in an experimental field with plant heights of 0.5, 0.9, 1.5, 2.0, and 2.5 m. The software was unable to identify planting failures smaller than 1.0 m when using pixel sizes larger than 3.5 cm. Even with the highest spatial resolution of 3.5 cm, the software could not detect planting failures between 0.5 and 1.0 m for sugarcane plants that were 0.9 m or taller.

3. Materials and Methods

This project comprises four activities outlined in Figure 1, illustrating the work breakdown structure (WBS) for developing a system to map gaps in sugarcane fields using RGB drone imagery. Below is a concise overview of each activity:

(1)

The initial activity involves capturing fields images, and following image capture, generating orthomosaics from the acquired images.

(2)

The second activity involves segmenting samples from the orthomosaic to train the convolutional neural network and generate models for mapping field gaps.

(3)

In the third phase, the optimal model will be utilized to identify field gaps within the orthomosaics.

(4)

Lastly, the fourth activity entails producing the mapping of field gaps, accompanied by a comprehensive report detailing the location of these gaps.

3.1. Taking Field Images and Generating Orthomosaics

This task involves generating an image database sourced from the sugarcane plantation, which will later serve as the foundation for mapping the locations of field gaps. For this purpose, a Drone DJI Phantom 4 Standard equipped with its 20 MP RGB camera was employed (Figure 2). The experimental fields for this endeavor were situated in the municipality of Tambaú, state of São Paulo, Brazil, at the following coordinates: latitude: −21.7023, longitude: −47.2814. The drones conducted flights over the crops, capturing images from an altitude of 60 m when the sugarcane plants were between three and five months old after sprouting. This altitude provided a GSD of 1.4 cm/pixel, which was sufficient to identify, select, and map the planting failures in a test image.

This task also involved creating orthomosaics (Figure 2). A total of 545 images were captured from the experimental field and stored in a database for orthomosaic construction using the OpenDroneMap 1.9.14 software [24]. The images were taken using the Pix4D (https://www.pix4d.com/) application, which allows for flight plan tracing and the adjustment of image overlap. In this instance, the overlap was set at 90% to facilitate the construction of the orthomosaic [25].

3.2. Training the YOLO Model

During the training phase of the convolutional neural network, it is crucial to have multiple input images of consistent dimensions, each containing samples of the objects that the neural network is designed to detect. These input images must have a resolution that is a multiple of 32 pixels. In this study, the input images were obtained by manually cropping sub-images from the orthomosaic. The chosen size for the input images was 416 × 416 pixels²; a standard dimension for such purposes [26,27]. The neural network in this study was tasked with identifying two classes: field gaps designated as class 0 and plants designated as class 1, as depicted in Figure 3.

Annex A includes the command utilized for neural network training. Subsequently, image samples and corresponding annotations for the two classes were employed to train the YOLOv5 neural network using the Google Colab platform [28,29] (Figure 4), negating the need to invest in computers equipped with GPUs for this endeavor. LabelImg software 1.8.6 [30,31] was utilized for class annotations (Figure 5).

3.3. Applying the YOLO Best Model

Upon completion of YOLOv5 training, executing the “detect” command with the best model file will proceed the identification of field gaps within the orthomosaics (Figure 6). Figure 7 illustrates field gaps, identified by a red rectangle. In this figure, the rectangles emphasize the field gap class to enhance visualization. Moreover, it omits the confidence index, which typically indicates the accuracy of object identification, the class name (whether field gap or plant), and any potential overlaps. Annex B encompasses the command utilized for field gap detection.

3.4. Generate Gaps Mapping

The fifth and final task in the field gap map process involves selecting the identified gaps based on a parameterized size measured in pixels². This approach is necessary because convolutional neural networks excel at identifying objects, regardless of their size in the image. Consequently, for planting failures, only empty spaces larger than 50 cm are considered to be commercially significant. The goal is to map only these substantial failures.

To address this, the mapping process is conducted in two stages. Initially, convolutional neural networks are used to identify planting failures of any size. In the second stage, commercially significant planting failures are selected. These failures will then be mapped using ImageJ digital image processing software.

Since YOLO identifies objects within rectangular bounding boxes, the selection of planting failures identified by YOLO must be refined using ImageJ based on the size of the area demarcated by YOLO. Typically, gaps exceeding 50 cm in length are classified as failures. To achieve the selection based on area, the width of the bounding box is multiplied by its length. However, these dimensions are in pixels, so a conversion is necessary to determine the length in centimeters.

In Figure 8, the red line represents the distance between cultivation rows, which is 150 cm for sugarcane. The red line, as shown in

Table 1. Lengths of lines in Figure 8.

Color	X	Y	Length Pixels	Area Pixels2	Length cm
Red	329.00	213.00	114.54		150.00
Orange	215.50	103.00	54.08	2595.84	70.82
Brown	190.00	73.00	48.00		62.86
Orange	154.50	191.50	61.01	4209.69	79.90
Blue	116.50	218.00	69.00		90.36
Orange	249.00	218.00	52.15	2816.10	68.29
Green	222.00	240.00	54.00		70.72
Orange	404.00	295.50	53.04	2917.73	69.46
Pink	377.50	321.50	55.01		72.04

, measures 114.54 pixels. This pixel length was obtained from ImageJ by drawing a line and selecting Analyze > Measure. The same method was applied to the lines in orange, brown, blue, green, and pink. Using the known measurement of 150 cm and the corresponding pixel length, an index was derived to convert all pixel lengths to centimeters, as shown in Table 1.

To differentiate between failure sizes to be recorded and those to be discarded, the area in pixels must be calculated. In Figure 8, an example using a subimage of Figure 16, where the rectangle in black marked with the number 3 is selected as a failure to be recorded, while the white-bordered rectangles should be discarded. Referring to Table 1, an area threshold of 4200 pixels² can be used to distinguish the black rectangle from the white-bordered ones (Figure 8a,b).

The index for converting pixel lengths to centimeters is calculated as follows:

Ind_pixel_cm = red line length/length in pixels = 1.3096

(1)

Using this index, the lengths in pixels can be converted to centimeters for the rectangles in Figure 8. In Figure 8c, the widths, highlighted in orange, are approximately the same size, ranging from 52 to 61 pixels. The following formula is used to calculate the area for a length of 50 cm in pixels:

length in pixels = length in cm/Ind_pixel_cm

(2)

and for a length of 50 cm:

length in pixels for 50 cm = 50/1.3096 = 38.18 pixels

To estimate the area of a failure with a length of 50 cm, the following formula is used:

Area in pixels² = avg_width × length_pixel

(3)

The average width of the failures is denoted as avg_width, assuming avg_width = 55.07 pixels.

Area for 50 cm length = 38.18 × 55.07 = 2102.57

Therefore, an area of 2100 pixels² will be used as a benchmark in this study for a length of 50 cm. To illustrate the entire process, an area value of 4200 pixels² will be used to demonstrate the exclusion of certain planting failures.

Once the planting failures have been selected using the ImageJ script, a report is generated detailing the location and size of each gap within the image. We utilized the freely available software ImageJ [23,32] to execute the selection process and generate the report. Figure 9 illustrates the steps for obtaining the field gap map, followed by a description of each step in the ImageJ script.

The initial step involves splitting the RGB image into three channels (Image > Color > Split Channels) [33], each containing the field gaps identified by YOLOv5, because ImageJ functions mostly work with 8 bit images. This process permits the selection of field gaps based on size, along with other operations such as field gap numbering (Figure 10).

In the next step, the red channel is used to identify field gaps based on a specified size, utilizing the built-in Analyze Particles plugin of ImageJ [34]. For this instance, the size parameter was set to 4200 pixels². This method was chosen because the neural network can sometimes identify smaller empty spaces that may not be relevant for commercial field gap analysis. This step helps to correct such potential misidentifications. Figure 11a illustrates the outcome of this procedure. Additionally, this analysis is set to generate a report in a table format, indicating the locations of the field gap map, as demonstrated in

Table 2. List of planting failures with sizes in pixels² and their location in the image.

	Area	X	Y	XM	YM	BX	BY	Width	Length
1	5670	533.72	21.60	533.53	21.72	468	0	132	45
2	5188	654.63	133.92	654.34	133.81	609	104	91	60
3	4255	117.99	192.41	117.77	192.16	81	161	73	61
4	5189	1053.96	379.69	1053.62	379.81	1013	348	82	64
5	4938	788.24	397.10	788.13	397.13	741	368	94	60
6	13590	752.44	661.61	751.91	661.73	642	608	218	123
7	4710	173.57	727.64	173.62	727.65	130	695	85	66
8	4781	1184.46	755.00	1184.29	754.98	1144	724	81	62
9	6976	51.97	803.14	51.43	802.98	4	766	97	74
10	4686	527.03	881.00	526.98	880.94	488	850	78	62
11	7368	646.00	905.74	646.18	905.66	584	876	124	60
12	4906	46.46	933.42	46.17	933.40	12	896	68	74
13	4946	392.00	935.42	391.36	935.03	353	901	77	67
14	5235	164.97	960.40	164.50	960.14	126	926	78	68
15	13881	568.30	1001.60	568.03	1001.53	488	958	160	88
16	9342	1006.56	1024.18	1006.84	1023.92	950	982	115	84
17	4435	35.58	1038.10	35.42	1037.84	0	1006	74	64
18	4580	586.93	1107.25	586.55	1107.28	550	1073	72	67
19	5792	122.02	1174.49	122.09	1174.15	76	1142	92	64
20	5598	228.96	1201.99	228.83	1201.77	184	1170	90	65
21	4410	806.00	1278.01	806.00	1277.93	760	1254	92	48

.

In Figure 11b, the gaps are counted and numbered using the Analyze Particles plugin [35] based on their position, following a horizontal arrangement from top to bottom and a vertical arrangement from left to right. These numbers correspond to the lines in the field gap map table (Table 2), where each line represents a gap with the same numbering, as seen in the image. This table serves to verify the accuracy of the identification system and can be used in sugarcane replanting operations.

Table 2 displays the identified gaps, with the gap numbers in the image aligning with the “Gap” column, allowing for easy association of data between the table and the numbered field gaps in the image (Figure 11b). The second column in Table 2 indicates the area of each gap in pixels². The “X” and “Y” columns represent the centroid coordinates of the field gap in the image, which will eventually be replaced with georeferenced data, enabling precise identification and replanting operations in these areas if required by farmers in a continuation of this work. The “XM” and “YM” columns denote the center of mass of each gap. The “BX” and “BY” coordinates specify the upper-left corner of the gap rectangle. The “Width” and “Length” columns provide the dimensions of each gap, respectively.

Figure 11c presents an inversion of the values from Figure 11b (Edit > Invert) [36], where the identified gaps are represented by maximum pixel values (255). Subsequently, subtracting Figure 11c from the original image highlights the field gaps and their corresponding numbers, as depicted in Figure 15. Since the field gaps are assigned maximum pixel values (255), subtracting these values (Process > Image Calculator) from the red, green, and blue channels [37] results in rectangles with zeroed-out pixels, effectively rendering them as black rectangles (Image > Color > Merge Channels) [38].

The subsequent step involves dividing the original image into channels (Figure 12). This process is designed to subtract the selected and numbered field gaps from each channel, as illustrated in Figure 13, and then merge (Image > Color > Merge Channels) these channels (Figure 14a for red, Figure 14b for green, and Figure 14c for blue) into a new image that displays the selected field map, depicted in Figure 14d. This newly merged image represents the field gap map report.

Figure 15a depicts the outcome of the selection process, while Figure 15b illustrates the output of YOLOv5′s detection of planting failures. In Figure 15c, the field gap selections are highlighted in black, and the rejected marks, deemed too small, are outlined using white rectangles (Analyze > Analyze Particles, with minimum size set to 4,200 pixels²). Figure 16 showcases the ultimate result of the field gap selection, providing improved visualization.

4. Results

The field gap map relies on two processes managed by two distinct files. The first file contains the YOLOv5 model, which is responsible for identifying all field gaps. Within this file are the neural network weights used for field gap identification. The training parameters included an input size of 416 × 416 pixels², a batch size of 16, and a planned duration of 300 epochs. However, training was stopped early at 210 epochs, after approximately 2 h, due to a lack of improvement in the validation metrics (early stopping). The final training loss achieved was 0.728, with a mean average precision (mAP) of 0.691.

The second file is the ImageJ script, which selects field gaps based on the minimum and maximum values specified during the Analyze Particles procedure. Additionally, it generates a table containing information such as the area and location of each selected field gap. After the selection process, the script assigns a unique number to each field gap according to the information provided in the table, enabling visualization of each field gap on a map and access to its corresponding data in the table.

Figure 17 shows the results of applying the best model to a larger image. The next step involves using an ImageJ script for selection. First, planting failures larger than 4200 pixels², which represent gaps greater than 100 cm, will be selected and mapped. Subsequently, failures larger than 2100 pixels², representing gaps larger than 50 cm, will be identified and mapped.

Figure 18 illustrates the output of the ImageJ script for failures exceeding 4200 pixels², identifying 53 planting failures with 3 instances of displacement errors, as shown in Figure 19. Based on the findings, the accuracy can be calculated by dividing these three misidentification issues by the total number of planting failures, resulting in an accuracy of 94.34%. The report detailing planting failures larger than 4200 pixels² is available in

Table A1. List of planting failures greater than 4200 pixels² with sizes given in pixels² and their location in Figure 18.

Gap	Area	X	Y	XM	YM	BX	BY	Width	Length
1	4611	1793.58	33.80	1793.19	33.72	1754	4	80	60
2	17,469	3097.97	66.13	3098.33	65.82	3016	10	172	122
3	5539	530.54	81.40	530.73	81.33	488	45	87	69
4	6368	1934.46	90.46	1933.78	90.39	1886	56	96	68
5	4593	663.66	143.40	663.44	143.07	626	112	76	62
6	7087	2312.32	144.99	2311.71	144.84	2246	118	132	54
7	6603	2428.00	156.45	2427.84	156.38	2384	118	88	76
8	4655	2264.53	213.66	2263.92	213.71	2230	180	70	68
9	7393	3152.99	233.01	3152.43	232.96	3096	199	113	67
10	6435	2534.51	260.00	2534.23	260.04	2477	232	115	56
11	6335	1081.59	326.70	1081.61	326.61	1036	292	93	70
12	7760	1956.02	429.10	1956.01	429.10	1910	385	92	87
13	4328	2566.47	433.49	2566.26	433.50	2530	402	72	62
14	4889	65.96	436.14	65.36	436.17	24	407	84	61
15	5499	1733.49	466.01	1733.56	466.00	1693	432	81	68
16	5392	3113.41	780.12	3112.97	780.20	3072	746	82	68
17	4343	2744.47	1517.40	2743.89	1516.89	2707	1486	73	62
18	5180	3118.36	1552.08	3118.36	1552.02	3072	1519	88	65
19	5661	2825.38	1671.02	2825.16	1670.73	2778	1640	94	62
20	4330	139.07	1743.97	138.78	1743.60	102	1714	76	60
21	5358	485.54	1770.50	485.36	1770.43	443	1738	87	65
22	6701	3087.52	1775.46	3087.28	1775.21	3036	1742	104	66
23	4419	3068.49	1862.01	3068.16	1861.87	3029	1834	79	56
24	5024	1294.60	1897.34	1294.24	1897.01	1256	1864	78	66
25	7979	938.11	1930.96	937.77	1930.83	886	1866	99	118
26	4256	1064.58	2010.36	1064.34	2010.25	1026	1980	78	58
27	4910	2564.82	2010.42	2564.47	2010.22	2513	1982	100	54
28	5489	389.88	2085.57	389.96	2085.55	340	2058	100	56
29	4231	1345.08	2121.36	1344.33	2121.21	1308	2092	74	58
30	6105	2073.64	2200.35	2073.68	2199.93	2022	2170	104	60
31	4698	1454.52	2275.03	1454.32	2275.00	1410	2248	89	55
32	9487	2326.71	2311.40	2327.03	2311.14	2258	2276	138	70
33	7766	110.06	2326.41	109.82	2326.24	60	2286	100	80
34	5116	769.07	2356.35	768.57	2356.16	730	2322	78	68
35	5340	1274.15	2426.99	1273.80	2426.79	1231	2396	87	62
36	7767	216.53	2521.59	216.26	2521.62	154	2490	124	64
37	5928	2694.50	2559.42	2694.32	2559.25	2648	2526	92	66
38	8423	2828.43	2582.59	2828.38	2582.58	2770	2546	116	74
39	7344	1926.12	2591.37	1926.24	2591.22	1857	2564	139	54
40	4274	1770.94	2642.68	1771.06	2642.95	1730	2616	82	54
41	4400	728.02	2673.00	727.90	2672.91	690	2644	76	58
42	4673	2180.98	2682.01	2180.42	2681.92	2142	2652	78	60
43	7462	1030.02	2700.60	1029.78	2700.49	984	2660	92	83
44	5112	3131.53	2709.25	3131.07	2708.94	3087	2680	89	58
45	5026	1494.50	2764.45	1493.52	2764.19	1456	2730	76	68
46	7834	464.09	2796.32	464.00	2796.03	416	2754	96	84
47	7612	1611.04	2906.01	1610.91	2905.92	1560	2868	102	76
48	5505	3120.99	2918.00	3120.50	2917.95	3068	2892	106	52
49	4228	216.39	2929.32	216.25	2929.25	182	2895	71	65
50	6036	1229.49	2973.78	1229.23	2973.79	1182	2942	96	64
51	5122	1904.05	3027.44	1903.69	3027.16	1862	2996	84	62
52	4612	301.06	3059.96	300.84	3059.81	265	3027	73	65
53	5271	760.01	3146.02	759.95	3146.10	716	3116	88	60

of Appendix A.

Figure 20 shows the detection of failures larger than 2100 pixels², resulting in the identification of 154 planting failures. In this second analysis, in addition to the displacement errors, three false negatives were identified (Figure 21). These false negatives, which are smaller than 4200 pixels², were only considered in the second run. The accuracy in this second analysis is the 6 errors (three misidentifications and three false negatives) divided by 154, resulting in a 96.1% accuracy. The report mapping failures larger than 2100 pixels² can be found in

Table A2. List of planting failures greater than 2100 pixels² with sizes given in pixels² and their location in Figure 20.

Gap	Area	X	Y	XM	YM	BX	BY	Width	Length
1	3041	897.42	22.01	896.81	22.14	862	0	70	44
2	3113	1153.01	20.01	1152.90	19.83	1114	0	78	40
3	4611	1793.58	33.80	1793.19	33.72	1754	4	80	60
4	2984	1927.35	29.59	1927.09	29.73	1890	8	72	44
5	17,469	3097.97	66.13	3098.33	65.82	3016	10	172	122
6	3711	1489.00	40.06	1488.51	40.08	1452	14	74	53
7	3481	1570.53	70.02	1570.05	69.97	1537	44	69	52
8	5539	530.54	81.40	530.73	81.33	488	45	87	69
9	3236	2186.58	73.42	2186.36	72.99	2156	46	62	54
10	6368	1934.46	90.46	1933.78	90.39	1886	56	96	68
11	3665	809.57	110.03	809.33	109.99	778	80	64	61
12	2499	2058.49	134.44	2058.03	133.78	2034	108	50	52
13	2717	1755.45	135.97	1755.27	136.04	1726	111	59	49
14	4593	663.66	143.40	663.44	143.07	626	112	76	62
15	7087	2312.32	144.99	2311.71	144.84	2246	118	132	54
16	6603	2428.00	156.45	2427.84	156.38	2384	118	88	76
17	2604	2689.48	167.02	2689.26	167.20	2666	138	46	58
18	4655	2264.53	213.66	2263.92	213.71	2230	180	70	68
19	7393	3152.99	233.01	3152.43	232.96	3096	199	113	67
20	2814	2048.46	248.36	2048.20	248.28	2021	222	55	52
21	6435	2534.51	260.00	2534.23	260.04	2477	232	115	56
22	4046	1685.36	324.04	1685.05	323.71	1652	290	66	66
23	6335	1081.59	326.70	1081.61	326.61	1036	292	93	70
24	3421	2511.57	330.02	2511.40	330.05	2484	298	56	65
25	3527	2271.99	332.01	2271.81	331.95	2238	306	68	52
26	2970	2920.02	359.52	2919.58	359.70	2892	332	56	56
27	3131	1509.48	371.50	1509.42	371.84	1481	344	57	56
28	3182	1835.45	394.02	1835.16	394.01	1808	364	54	60
29	3308	2406.96	393.45	2406.97	393.31	2376	364	60	58
30	3482	2200.01	405.00	2199.74	405.02	2170	374	60	62
31	7760	1956.02	429.10	1956.01	429.10	1910	385	92	87
32	4328	2566.47	433.49	2566.26	433.50	2530	402	72	62
33	4889	65.96	436.14	65.36	436.17	24	407	84	61
34	5499	1733.49	466.01	1733.56	466.00	1693	432	81	68
35	3812	1543.11	503.93	1542.55	503.88	1509	476	70	56
36	3990	2763.02	533.01	2762.82	532.93	2726	506	74	54
37	2644	2940.99	606.45	2940.81	606.37	2920	574	42	64
38	2704	687.46	606.46	686.84	606.36	660	580	54	52
39	3697	1822.44	608.35	1821.94	608.20	1787	580	69	56
40	2376	1924.48	638.93	1924.43	638.75	1903	610	43	58
41	2438	661.52	698.02	661.18	697.74	638	672	48	52
42	2357	2518.91	769.40	2518.88	769.27	2495	742	47	54
43	5392	3113.41	780.12	3112.97	780.20	3072	746	82	68
44	3182	1126.50	892.98	1126.11	893.05	1099	864	55	58
45	2455	1373.51	892.55	1373.32	892.77	1352	864	44	58
46	3376	2564.88	901.16	2564.67	901.30	2534	872	61	68
47	2615	926.07	923.85	925.88	923.83	900	898	52	52
48	3248	1222.54	926.98	1222.48	927.17	1190	902	66	50
49	2691	2530.54	989.50	2530.30	989.43	2506	962	50	55
50	3842	453.43	1070.61	452.90	1070.49	422	1038	61	70
51	3352	1645.98	1098.00	1645.44	1097.58	1616	1070	60	56
52	3027	923.48	1161.55	923.15	1161.55	896	1134	55	56
53	2717	2406.00	1166.04	2406.00	1166.19	2378	1140	56	52
54	2862	3170.04	1354.42	3169.87	1354.54	3144	1326	52	56
55	3013	1762.07	1371.97	1761.49	1372.03	1733	1346	59	52
56	3199	3147.55	1450.12	3147.28	1450.00	3116	1422	63	57
57	2555	2846.43	1466.01	2846.23	1465.74	2822	1438	48	56
58	3064	2413.47	1497.95	2413.03	1498.02	2383	1470	60	54
59	2915	2132.45	1502.52	2131.98	1502.43	2104	1476	56	54
60	4343	2744.47	1517.40	2743.89	1516.89	2707	1486	73	62
61	5180	3118.36	1552.08	3118.36	1552.02	3072	1519	88	65
62	3831	1746.49	1564.99	1746.24	1565.04	1710	1538	73	54
63	3660	100.98	1597.57	100.80	1597.31	70	1568	62	60
64	2548	2646.46	1594.04	2646.48	1594.05	2621	1568	51	53
65	3021	2328.00	1596.01	2327.49	1595.96	2300	1569	56	55
66	3458	710.02	1650.00	709.84	1649.78	678	1622	64	56
67	2150	227.58	1664.96	227.29	1665.13	206	1640	44	50
68	5661	2825.38	1671.02	2825.16	1670.73	2778	1640	94	62
69	2307	19.49	1692.42	19.44	1692.62	0	1662	40	61
70	3031	369.55	1721.38	369.08	1721.32	340	1694	61	54
71	3486	890.46	1740.67	890.08	1740.38	862	1710	57	62
72	4330	139.07	1743.97	138.78	1743.60	102	1714	76	60
73	5358	485.54	1770.50	485.36	1770.43	443	1738	87	65
74	6701	3087.52	1775.46	3087.28	1775.21	3036	1742	104	66
75	2970	2841.46	1799.00	2841.44	1799.04	2815	1770	53	58
76	3773	1098.03	1828.46	1097.95	1828.29	1066	1798	64	60
77	3048	2657.08	1844.31	2656.91	1844.04	2632	1812	50	64
78	4419	3068.49	1862.01	3068.16	1861.87	3029	1834	79	56
79	5024	1294.60	1897.34	1294.24	1897.01	1256	1864	78	66
80	7979	938.11	1930.96	937.77	1930.83	886	1866	99	118
81	3082	2840.97	1901.55	2840.78	1901.55	2812	1874	58	56
82	3076	544.00	1914.48	543.72	1914.51	516	1886	56	56
83	3833	1065.55	1930.64	1065.46	1930.92	1032	1902	68	58
84	3294	2842.56	2007.64	2842.40	2007.64	2818	1974	50	69
85	4256	1064.58	2010.36	1064.34	2010.25	1026	1980	78	58
86	4910	2564.82	2010.42	2564.47	2010.22	2513	1982	100	54
87	2804	1355.51	2026.51	1355.38	2026.59	1329	2000	53	53
88	3972	122.45	2087.67	122.01	2087.74	88	2058	68	60
89	5489	389.88	2085.57	389.96	2085.55	340	2058	100	56
90	2648	714.96	2101.60	714.82	2101.75	690	2074	50	56
91	4231	1345.08	2121.36	1344.33	2121.21	1308	2092	74	58
92	3806	1698.52	2129.16	1698.47	2129.09	1667	2098	63	62
93	3847	241.50	2130.39	240.97	2130.53	208	2100	66	60
94	3450	857.51	2153.38	856.91	2153.32	825	2126	65	54
95	3020	352.54	2181.41	352.03	2181.39	326	2152	53	58
96	4035	640.95	2195.64	640.32	2195.63	610	2162	62	69
97	6105	2073.64	2200.35	2073.68	2199.93	2022	2170	104	60
98	3632	1884.97	2227.56	1884.91	2227.65	1852	2200	66	56
99	4698	1454.52	2275.03	1454.32	2275.00	1410	2248	89	55
100	9487	2326.71	2311.40	2327.03	2311.14	2258	2276	138	70
101	3525	660.97	2310.49	660.88	2310.57	628	2282	65	56
102	7766	110.06	2326.41	109.82	2326.24	60	2286	100	80
103	4157	1558.17	2313.96	1557.94	2313.73	1521	2286	75	56
104	5116	769.07	2356.35	768.57	2356.16	730	2322	78	68
105	3613	1937.10	2363.45	1937.15	2363.34	1903	2336	70	54
106	3776	1143.98	2383.53	1143.52	2383.76	1112	2354	64	61
107	5340	1274.15	2426.99	1273.80	2426.79	1231	2396	87	62
108	2650	811.01	2480.48	810.54	2480.49	785	2454	53	52
109	3601	28.49	2517.60	28.10	2517.77	0	2486	58	64
110	7767	216.53	2521.59	216.26	2521.62	154	2490	124	64
111	4087	2080.01	2532.39	2079.82	2532.30	2042	2504	76	56
112	3131	1485.50	2533.48	1485.24	2533.37	1458	2505	56	57
113	3472	3033.02	2549.50	3032.37	2549.23	2998	2524	70	52
114	5928	2694.50	2559.42	2694.32	2559.25	2648	2526	92	66
115	3306	2491.95	2572.41	2491.65	2572.32	2462	2539	60	61
116	8423	2828.43	2582.59	2828.38	2582.58	2770	2546	116	74
117	4069	232.69	2589.20	232.89	2589.25	194	2562	80	57
118	7344	1926.12	2591.37	1926.24	2591.22	1857	2564	139	54
119	3955	574.99	2624.00	574.71	2624.12	536	2598	78	52
120	4274	1770.94	2642.68	1771.06	2642.95	1730	2616	82	54
121	4400	728.02	2673.00	727.90	2672.91	690	2644	76	58
122	2658	1854.98	2673.53	1854.82	2673.64	1828	2648	54	54
123	4673	2180.98	2682.01	2180.42	2681.92	2142	2652	78	60
124	7462	1030.02	2700.60	1029.78	2700.49	984	2660	92	83
125	3236	2479.45	2688.58	2479.11	2688.67	2449	2662	61	54
126	5112	3131.53	2709.25	3131.07	2708.94	3087	2680	89	58
127	5026	1494.50	2764.45	1493.52	2764.19	1456	2730	76	68
128	7834	464.09	2796.32	464.00	2796.03	416	2754	96	84
129	3413	1754.54	2847.98	1754.12	2847.88	1726	2818	58	60
130	3015	349.03	2862.97	348.82	2862.94	322	2834	54	58
131	2355	2387.57	2869.96	2387.36	2869.52	2364	2844	48	52
132	2489	2721.54	2885.48	2721.30	2885.37	2698	2858	48	54
133	7612	1611.04	2906.01	1610.91	2905.92	1560	2868	102	76
134	3281	1366.50	2915.48	1366.13	2915.30	1332	2890	68	50
135	5505	3120.99	2918.00	3120.50	2917.95	3068	2892	106	52
136	4228	216.39	2929.32	216.25	2929.25	182	2895	71	65
137	3107	2545.00	2934.01	2544.88	2934.18	2514	2908	62	52
138	3077	2616.55	2952.97	2616.45	2952.96	2588	2926	58	54
139	6036	1229.49	2973.78	1229.23	2973.79	1182	2942	96	64
140	3676	2725.50	3007.13	2725.44	3007.20	2694	2976	62	64
141	3751	176.23	3020.24	176.21	3020.35	143	2990	68	64
142	2846	2798.51	3020.00	2798.19	3020.03	2770	2994	57	51
143	5122	1904.05	3027.44	1903.69	3027.16	1862	2996	84	62
144	2644	2508.46	3024.00	2508.39	3024.04	2483	2998	51	52
145	4612	301.06	3059.96	300.84	3059.81	265	3027	73	65
146	2742	789.01	3067.12	788.95	3067.12	765	3037	49	59
147	2588	579.55	3070.01	579.33	3070.04	554	3042	52	54
148	3704	1472.13	3077.96	1471.79	3077.59	1438	3050	70	56
149	5271	760.01	3146.02	759.95	3146.10	716	3116	88	60
150	3972	554.44	3149.65	554.11	3149.72	520	3120	68	60
151	3892	1096.00	3155.35	1096.08	3155.36	1062	3126	68	58
152	2697	3186.94	3168.42	3186.70	3168.64	3162	3135	47	65
153	3911	1470.39	3170.93	1469.95	3170.58	1434	3140	68	62
154	2408	3008.06	3186.51	3008.03	3186.49	2980	3164	58	45

of Appendix B.

5. Discussion and Conclusions

Both analyses, for gaps larger than 4200 pixels² and those larger than 2100 pixels², outperformed those in [14]. The developed application shows promise for effectively identifying small field gaps between 50 and 100 cm. The YOLOv5 neural network successfully identified nearly all the field gaps, allowing users to customize the threshold for gap size consideration and choose whether to discard smaller gaps based on their specific needs.

Field gaps present a significant challenge to sugarcane productivity, especially given the crop’s semi-perennial nature. These gaps not only affect the yield of the current year but also have repercussions on productivity in subsequent years. The identification and mapping of field gaps play a crucial role in informing replanting strategies and estimating productivity.

Some studies have focused on identifying and mapping field gaps, but they often struggle to detect small gaps due to spatial resolution limitations and occlusions from sugarcane leaves [11,12]. In our solution, using GSD imagery with a resolution of 1.4 cm per pixel, we successfully mapped almost all the small gaps of 50 cm or larger in an image of 2309 × 2309 pixels. Since ImageJ selects particles based on area, the 50 cm gaps were converted to an area of 2100 pixels².

The application presented in this work has the potential to be an innovative solution by integrating computer vision, machine learning, convolutional neural networks, and digital image processing. Recognizing objects in images is challenging due to variations in scale, as objects may appear at different distances, yet their sizes still need to be identified. For instance, when detecting planting failures, objects of various sizes can be considered failures, but economically significant failures are typically those longer than 50 cm. Convolutional neural networks are not inherently suited for this task, so using a digital image processing tool like ImageJ enables the accurate selection of empty spaces that truly qualify as failures. This approach also offers flexibility for producers to tailor the identification process, accommodating factors such as failures caused by improperly calibrated harvesting machines and other equipment.

Since occlusions from leaves or weeds reduce the visible space between tillers, thereby concealing planting failures, we plan to treat these spaces as indicators of planting failures in future research. This approach will enhance the accuracy of identifying and mapping planting failures.