Application of YOLO-based deep learning and UAV imagery in hybrid maize seed production

ABSTRACT

In maize breeding programs, emasculation is a critical step in the production of hybrid seeds, occurring prior to anthesis. Therefore, an accurate identification of tassels is paramount for the efficient execution of detasseling. However, tassel detection poses significant challenges due to environmental variability in production fields, the spatial resolution of aerial images, lighting conditions, and the phenotypic variability of the plants. Traditional methods are also predisposed to subjectivity and detasseling errors. This study proposes a solution to these challenges by integrating a remote sensing platform, utilizing an unmanned aerial vehicle (UAV), with a deep learning (DL) algorithm to enhance the efficiency of maize tassel detection. The UAV captured the red-green-blue (RGB) images spectral region at an altitude of 20 m, and the YOLOv4 algorithm was employed to develop the tassel detection model. The image database was constructed using preprocessing strategies to augment the dataset. The detection model based on the YOLOv4 architecture demonstrated satisfactory results, with an overall accuracy of 0.89, ranging from 0.84 to 0.92, depending on the levels of detasseling errors and the morphological characteristics of maize tassels. This study underscores the applicability of YOLOv4 in detecting plants with detasseling errors in seed production fields, thereby contributing to the purity and quality of hybrid seeds.

YOLOv4; aerial image; detection; tassel

Introduction

Approximately 95 % of the global maize (Zea mays L.) producing regions, encompassing nearly all cultivated areas, cultivate hybrid maize varieties (Song et al., 2023). In Brazil, single-cross hybrids predominate, accounting for over 50.19 % of the maize market, followed by triple-cross hybrids (1.93 %), double-cross hybrids (1.16 %), and open-pollinated varieties (2.70 %) (Pereira Filho and Borghi, 2022). The production of hybrid seeds necessitates the presence of specific environmental conditions, including optimal temperature and humidity levels, in addition to specialized technical expertise for large-scale operations (Shi et al., 2020).

The genetic purity of seeds is of paramount importance in maize seed production, with the size and number of tassel branches exerting a significant influence on this quality (Lambert and Johnson, 1978). The meticulous monitoring of maize tassels and pollination is imperative for the successful production of hybrid seeds, with emasculation (detasseling) emerging as a pivotal process. Manual detasseling is characterized by delays, high labor intensity, and increased costs, while mechanical detasseling, although more efficient, faces challenges related to terrain and plant morphology (Herr et al., 2023; Jia et al., 2024; Song et al., 2023).

Accurate and cost-effective identification of maize tassels during the flowering stage is imperative for enhancing crop yield and quality (Niu et al., 2024). The development of computer vision (CV) and deep learning (DL) methodologies for detecting emasculation failures in seed production fields is essential for the advancement of intelligent agriculture.

Recent advancements in CV and artificial intelligence (AI) have led to significant progress in the field of object detection using images captured by unmanned aerial vehicles (UAVs). These advancements have resulted in the ability to detect objects with high spatiotemporal resolution (Liu et al., 2022; Lu et al., 2024). DL models, including neural networks, have proven highly effective in detecting and quantifying objects within images (Sirisha et al., 2023). The YOLO (You Only Look Once) object detection model partitions images into grids to identify objects within each cell, offering a compact model size and fast processing speed (Lu et al., 2024). YOLO has been widely applied to tasks such as tassel detection, plant counting, and disease identification (Zhao et al., 2023; Zhao et al., 2021).

The objective of this study is twofold: first, to train a maize tassel detection model using YOLOv4 and UAV RGB images, and second, to evaluate the model’s robustness in environments exhibiting varying tassel morphology and detasseling success.

Materials and Methods

Study site and experimental design

The present study was conducted in the municipality of Ijaci, Minas Gerais state, Brazil (21°9’51.94” S, 44°55’6.17” W, altitude 845 m) (Figure 1). The region exhibits an average temperature of 21 °C and an annual precipitation of 1,350 mm, defining a subtropical climate with dry winters and hot summers (Cwa) (Alvares et al., 2013).

Figure 1
– Location of the field: municipality of Ijaci, Minas Gerais state, Brazil.

The data was collected during the 2023/2024 agricultural season, targeting maize (Z. mays) cultivation. The field was subdivided into 150 plots, with each experimental unit consisting of four rows spaced 0.5 m apart and 5 m in length, resulting in a total area of 10 m² per plot. The experimental design employed was a randomized complete block design (RCBD), encompassing 25 treatments, including five maize hybrids and five levels of simulated detasseling failures in the field, with these treatments replicated six times and arranged in an RCBD.

On 22 Oct 2023, the cultivation of maize was initiated through conventional and mechanized methods, encompassing five distinct hybrids: P4285 VYHR, B2801 PWU, P3016 VYHR, B2782 PWU, and LG36799 VIP3. The development of a robust and generalized system capable of field inference from aerial imagery is imperative, and training a YOLOv4 model with the diverse characteristics of maize tassels is essential for this endeavor. Tassel appearance can vary significantly due to factors such as genetic differences among hybrids, and environmental influences such as lighting, shadows, and canopy density.

Incorporating these variations during training enables the model to effectively adapt to real-world conditions, thereby ensuring accurate detection across heterogeneous datasets. This approach reduces overfitting and enhances the model’s generalization ability, improving its reliability in precision agriculture. As a result, it facilitates scalable and accurate tassel detection for various field scenarios (Niu et al., 2024; Song et al., 2023).

Fungicides, insecticides, and herbicides were applied as needed in the experimental area, using a remotely piloted aircraft, model DJI Agras T10 (SZ DJI Technology Co.), with a 10-liter tank capacity, a flow rate of 12 L ha^–1, and an application height of 3 m above the canopy.

Data acquisition and image processing

The simulation of detasseling failures was conducted in the field according to five contamination levels: i) 1 % of the plants in the plot with tassels, ii) 2 % of the plants with tassels, iii) 3 % of the plants with tassels, iv) 4 % of the plants with tassels, and v) 5 % of the plants with tassels. The detasseling process was executed manually, entailing the tassel’s removal prior to the flag leaf’s full expansion. The imagery was captured through the utilization of an UAV, precisely, the Phantom 4 model (SZ DJI Technology Co.), which was equipped with an RGB sensor (model FC330, DJI Technology Co.). The images were then employed to generate an orthomosaic and construct a database.

The aerial survey was conducted on 26 Dec 2023 coinciding with the simulated detasseling failures in the maize plants. Flights were conducted between 11h00 and 13h00 local time to minimize the impact of incident radiation and shadows on the imagery. The flight plan was meticulously designed using Pix4D Capture software (Pix4D SA, version 4.13.1), incorporating 80 % front and side overlap, a flight altitude of 20 m, an RGB camera oriented perpendicularly to the ground, and a ground sampling distance of 0.55 cm per pixel. A total of 196 aerial images were captured during the mission.

Eight ground control points were distributed uniformly across the field to obtain precise geographical references (Ferraz et al., 2024) and enhance the detail and accuracy of crop information extraction. The geographic coordinates of the points were collected using GPS equipment with Real-Time Kinematic (RTK) signal correction (model D-RTK 2, SZ DJI Technology Co.). The orthomosaic generation was facilitated by Pix4D Mapper software (Pix4D SA, version 4.5), following a processing workflow (Figure 2) that entailed image alignment, GCP marking, realignment, dense point cloud construction, texture generation, and orthomosaic creation.

Figure 2
– Image processing workflow in Pix4D Mapper software. UAV = unmanned aerial vehicle; RGB = red-green-blue spectral images; GCP = ground control point.

Plot extraction from the orthomosaic

Images corresponding to each experimental unit were extracted from the orthomosaic using QGIS software (QGIS Development Team, version 3.22.15). The Plot Boundary plugin (version 0.1.2) within QGIS was employed to create the plots, which were applied as a mask layer during the raster file cropping process. The raster file subsets, derived from the orthomosaic, were then used as input for the YOLOv4 model inference detect and count maize tassels accurately.

Annotation setup and database configuration

A database was initially created, consisting of raw images captured by the RGB sensor. The dataset was divided into a training set (80 % of the total dataset) and a testing set (20 %), with annotations made solely for a single class of interest. The training dataset included 9,068 images, while the test dataset contained 2,267 images. The “tassel” class annotations included 45,340 instances in the training set and 11,335 instances in the test. This partitioning was designed to ensure a robust and representative sampling for model development and evaluation. The number of image samples is crucial in developing effective image detectors. While datasets may contain hundreds to several thousand samples for training, the variability may still be insufficient to build an accurate model.

Image augmentation techniques were employed during training on the Roboflow platform. These techniques included cropping, brightness adjustment, noise addition, exposure modification, 90° rotation, mosaic augmentation, scaling, translation, random shearing, mix up augmentation, and random changes in hue, saturation, and value, as well as horizontal and vertical flipping (Figure 3A-I). These augmentation functions were applied at the image level, with an average of five tassel annotations per image, enhancing model robustness and mitigating overfitting issues.

Figure 3
– Illustration of the image augmentation pipeline for training the YOLOv4 detection algorithm. A) From the original image, the following augmentation techniques were applied: B) vertical flip, C) value adjustment, D) hue shift, E) brightness adjustment, F) saturation adjustment, G) rotation, H) darkening, and I) mosaic transformation.

Annotations were made from the RGB image database, and the coordinates of the bounding boxes responsible for locating the objects of interest were calculated. A bounding box presents the necessary information about the object for the model training process. Similarly, the detection results will be displayed in the image as bounding boxes (Diwan et al., 2023).

YOLO algorithm hyperparameters

The experiment was conducted in a computational environment with an Intel Core i7-9700KF processor (2.6 GHz), 32 GB of RAM, a NVIDIA GeForce RTX 2060 GPU (6 GB), and the Windows 10 operating system. The training and testing of the tassel detection model were performed within the Google Collaboratory virtual environment (Google LLC), utilizing a T4 GPU hardware accelerator and 15 GB of RAM. The implementation of the YOLOv4 algorithm was carried out in Python Language (Python Software Foundation, version 3.11) (Table 1), and the neural network was trained based on the number of iterations required for convergence.

Table 1
– Hyperparameters and specifications of the YOLOv4 detection algorithm of objects.

Hyperparameters	Detailing
Detect algorithm	YOLOv4
Feature extractor	CSPDarknet-53
Input size	896 × 896
Number of parameters	6.4 × 107
Learning rate	1 × 10–3
Batch size	64
Number of iterations	8000
Code repository	github.com/AlexeyAB/darknet

Statistical metrics

The performance of the YOLOv4 algorithm in detecting tassels in maize crops was evaluated using commonly employed statistical metrics (Lu et al., 2022; Sirisha et al., 2023). A model prediction can be classified into four prediction cases: i) true positive (TP), ii) false positive (FP), iii) true negative (TN), and iv) false negative (FN) (Lu et al., 2024).

Precision is defined as the proportion of TPs out of all detections (TPs and FPs) (Eq. 1):

$\text{Precision}=\frac{TP}{FP+TP}$ (1)

Recall is defined as the proportion of TPs out of the total number of relevant objects (TPs and FNs) (Eq. 2):

$\text{Recall}=\frac{TP}{FN+TP}$ (2)

F1-score is the harmonic mean of Precision and Recall (Eq. 3):

$\text{F1}-\text{score}=\frac{2\times \text{Precision}\times \text{Recall}}{\text{Precision}+\text{Recall}}$ (3)

Validation was performed by comparing the validation sample with the reference data from the simulated tassel failure levels conducted in the field. The detection performance and quality assessment used a confusion matrix and metrics, including overall accuracy, Precision, Recall, and F1-score. All statistical analyses throughout this study were executed using Python Language (Python Software Foundation, version 3.11) and the R Statistical Computational Environment (R Development Core Team, version 4.2.3).

Results

The evaluation metrics for the YOLOv4 object detection model, including overall mean Precision, Recall, and F1-score, are illustrated in Figure 4. The parameter values analyzed indicate satisfactory performance detecting simulated tassel detachment failures in the field. The YOLOv4 algorithm produced favorable results, reinforcing its potential application in maize breeding programs.

Figure 4
– Overall Precision, Recall, and F1-score for the “tassel” class at different confidence thresholds of the YOLOv4 algorithm detection on the test dataset.

The pattern of Precision, Recall, and F1-score metrics demonstrates the model’s performance across varying confidence thresholds. The final Precision and Recall values were calculated using a predefined confidence threshold (c = 0.25), achieving 0.88 and 0.85, respectively.

The model’s performance in object detection was validated through a confusion matrix analysis (Table 2) for the YOLOv4 algorithm. As previously highlighted by the satisfactory general performance metrics, minimal errors were observed between the classified thematic levels in the confusion matrix. Specifically, levels 2 and 5 exhibited a higher degree of confusion; the images corresponding to level 2 plots demonstrated some pixel distortion effects in the orthomosaic, which may have resulted in under- and over-detection of the target object. Level 5 was influenced by the structure of the male reproductive organ of the B2801 PWU hybrid, which also resulted in increased detection errors. However, these errors were minimal and did not compromise the statistical results of the detection performance for the study targets.

Table 2
– Confusion matrix for the “tassel” class across five tasseling levels in maize, detected by the YOLOv4 object detector.

	Tasseling level	Reference
		1		2		3		4		5
Prediction	1	21		1		0		0		1
	2	4		21		2		4		1
	3	0		2		25		1		2
	4	0		0		1		22		4
	5	1		0		0		0		18

The confusion matrix analysis for the detection model across the five tassel detachment levels is shown in Figure 5. Level 2 exhibited low Precision (0.66) and F1-score (0.75) values, attributable to distortion effects from photogrammetric processing of aerial images and the absence of RGB camera brightness calibration. Similarly, level 5 exhibited a low Recall value (0.69), attributable to model inference errors, predominantly FN detections, which exerted a more substantial influence on the prediction of Precision (0.95).

Figure 5
– Summary of the confusion matrix for the five tasseling levels in maize, detected by the YOLOv4 detection algorithm of objects.

The B2801 PWU hybrid’s tassel is distinguished by a slender rachis and a limited number of branches, exhibiting a light coloration that can hinder identification and lead to inaccuracies in quantifying the male organ. These results in the non-detection of more authentic objects and reduced Recall value (Figure 6). In contrast, the YOLOv4 model exhibited superior accuracy in detecting tassels for the other maize hybrids. Notably, the P4285 VYHR hybrid exhibited distinctive tassel morphological traits, including a significant number of branches and an orange-hued rachis, contributing to the model’s detection performance.

Figure 6
– YOLOv4 model performance by maize hybrid.

The efficacy of the YOLOv4 model in detecting maize tassels was assessed by a series of three randomized aerial images captured using an RGB camera mounted on a UAV. This approach was employed to ascertain the model’s capacity to identify the target in a field setting. Each bounding box center corresponds to a specific maize tassel, as illustrated in Figure 7A-E. Following the implementation of the tassel detector, the bounding boxes delineated in pink represent the detected tassels, with the object class designated as “tassel” and the confidence level at the time of inference displayed at the upper portion of each bounding box.

Figure 7
– Original images captured by unmanned aerial vehicle of maize tassels after detection by the YOLOv4 detection algorithm. Maize hybrids: A) P4285 VYHR, B) B2801 PWU, C) P3016 VYHR, D) B2782 PWU, and E) LG36799 VIP3.

The tassel identifiers for five maize hybrids (Figure 7A-E), P4285 VYHR, B2801 PWU, P3016 VYHR, B2782 PWU, and LG36799 VIP3, were exemplified using the YOLOv4 algorithm on random UAV images.

The method developed for maize tassel detection using CV was evaluated by comparing the detections obtained by the YOLOv4 algorithm with simulated tassel detachment failures. The model demonstrated its capacity for detection and inference based on images collected 20 m above the canopy. Specifically, the model can detect the number of plants with tassels (Figure 8A-E).

Figure 8
– Detection model inference from the orthomosaic for the five levels of tasseling deficiencies. A) Level 1, B) 2, C) 3, D) 4, and E) 5.

The model’s reduced accuracy in detecting maize tassels is attributable to the smaller number of labels resulting from the inherent subjectivity generated during the tassel annotation process for the B2801 PWU hybrid. The challenge is further compounded by the model’s inability to differentiate between tassels and the green color of the leaves due to their morphological characteristics, including size, shape, thickness, and overall similarity.

Discussion

Recent advancements and updates to the YOLOv4 algorithm have demonstrated the potential to deliver enhanced results while reducing computational costs (Falahat and Karami, 2023; Kumar et al., 2021; Song et al., 2023; Wang et al., 2023). In this study, model performance metrics decreased with higher confidence thresholds, a pattern consistent with the findings of Ammar et al. (2021), who also noted the superiority of the YOLOv4 algorithm over other networks.

The evaluation of the model performance through metrics is imperative, however, these metrics can vary according to the confidence threshold employed. Reducing this threshold results in models with higher positive predictions, consequently increasing Recall values. Conversely, increasing the confidence threshold promotes more conservative detections, potentially increasing precision due to more positive predictions (Rodríguez-Lira et al., 2024). For applications requiring high sensitivity, a lower confidence threshold is preferable. Conversely, scenarios where FPs are heavily penalized demand higher thresholds (Bazame et al., 2023).

Photogrammetric processing is susceptible to pixel distortions; however, camera contrast and calibration have been demonstrated to increase the number of homologous points between overlapping images, reducing reprojection error by approximately 10 %. This, in turn, results in a more representative orthomosaic of the area (Burdziakowski and Bobkowska, 2021). Furthermore, the inference errors could be attributed to a slightly imbalanced class distribution (Herrera et al., 2024), as indicated by the imbalance found at level 5 of the present study.

The statistical averages for accuracy (0.89), Precision (0.83), Recall (0.82), and F1-score (0.82) of the YOLOv4 object detector were consistent with the results of studies using YOLOv4 in agriculture (Ammar et al., 2021; Yin et al., 2023; Zhang et al., 2023). The F1-score, which harmoniously combines Precision and Recall, serves as a pivotal indicator for identifying the optimal confidence threshold for tassel detection across different maize hybrids. This harmonic metric provides a balanced evaluation of the YOLOv4 model’s performance.

Errors in tassel detection were observed, and they were attributed to several factors, which included tilted, pixelated images, distortions, and plant movement during the aerial survey. A potential solution for correcting the orthomosaic is to use image stitching techniques, recommended for dynamic captures at low altitudes and misaligned images (Zhang et al., 2022).

In order to address the challenges associated with plant phenotyping in complex field conditions, DL methods have been increasingly used, which possess the capacity to discern salient features from substantial image data (Yu et al., 2022). Nevertheless, it is imperative to acknowledge the challenges associated with the deployment of these models on images captured from elevated perspectives. This poses a dilemma wherein achieving a balance between precision, computational complexity, and the detection of multiple targets becomes arduous (Niu et al., 2024). Furthermore, factors such as the accuracy of simulated tassel detachment levels and the spatial resolution of aerial images have been demonstrated to influence detection accuracy (Herrera et al., 2024; Li et al., 2023).

The integration of object detection systems and aerial imagery in agriculture has led to a variety of applications, including the detection of different tree species, the counting of plants, and the detection of leaf diseases, among others. Various studies have utilized image processing and DL methods to count and detect fruits and flowers (Herrera et al., 2024).

The present study demonstrated the potential and applicability of integrating UAV-based aerial imagery with deep machine learning to detect tassels in maize plants that have not undergone detasseling. Implementing the YOLOv4 algorithm for tassel detection yielded consistent and satisfactory results in terms of statistical metrics, further underscoring its advantages over traditional methods. The tassel detection model attained an overall accuracy of 0.89, with variations ranging from 0.84 to 0.92, contingent upon the levels of detasseling failures and the morphological characteristics of the maize tassels.

This study’s findings suggest that utilizing YOLOv4 in hybrid maize seed production fields is a viable approach to ensure the preservation of genetic purity and the generation of high-quality seeds. This recommendation is aimed at adverting self-pollination, a process that can compromise seed integrity. Integrating a maize tassel detection model within the study framework is instrumental in identifying potential detasseling failures that may occur in female parent rows. This capacity is enabled through the employment of advanced techniques for aerial surveys, which facilitate a comprehensive acquisition of data in seed production fields.

Acknowledgments

The authors are especially grateful to the Fundação de Amparo à Pesquisa do Estado de Minas Gerais (FAPEMIG), the Conselho Nacional de Desenvolvimento Científico e Tecnológico (CNPq), the Coordenação de Aperfeiçoamento de Pessoal de Nível Superior (CAPES), and the Universidade Federal de Lavras (UFLA) for all support for the conduction of this research. The FAPEMIG funded this research through project 11680.

References

• Alvares CA, Stape JL, Sentelhas PC, Gonçalves JLM, Sparovek G. 2013. Köppen's climate classification map for Brazil. Meteorologische Zeitschrift 22: 711-728. http://dx.doi.org/10.1127/0941-2948/2013/0507
  » http://dx.doi.org/10.1127/0941-2948/2013/0507
• Ammar A , Koubaa A , Benjdira B . 2021. Deep-Learning-based Automated Palm Tree Counting and Geolocation in Large Farms from Aerial Geotagged Images. Agronomy 11: 1458. http://dx.doi.org/10.3390/agronomy11081458
  » http://dx.doi.org/10.3390/agronomy11081458
• Bazame HC, Molin JP, Althoff D, Martello M. 2023. Detection of coffee fruits on tree branches using computer vision. Scientia Agricola 80: e20220064. http://dx.doi.org/10.1590/1678-992x-2022-0064
  » http://dx.doi.org/10.1590/1678-992x-2022-0064
• Burdziakowski P, Bobkowska K. 2021. UAV Photogrammetry under poor lighting conditions: accuracy considerations. Sensors 21: 3531. http://dx.doi.org/10.3390/s21103531
  » http://dx.doi.org/10.3390/s21103531
• Diwan T, Anirudh G, Tembhurne JV. 2023. Object detection using YOLO: challenges, architectural successors, datasets and applications. Multimedia Tools and Applications 82: 9243-9275. http://dx.doi.org/10.1007/s11042-022-13644-y
  » http://dx.doi.org/10.1007/s11042-022-13644-y
• Falahat S, Karami A. 2023. Maize tassel detection and counting using a YOLOv5-based model. Multimedia Tools and Applications 82: 19521-19538. http://dx.doi.org/10.1007/s11042-022-14309-6
  » http://dx.doi.org/10.1007/s11042-022-14309-6
• Ferraz MAJ, Barboza TOC, Arantes PS, von Pinho RG, Santos AF. 2024. Integrating satellite and UAV technologies for maize plant height estimation using advanced machine learning. AgriEngineering 6: 20-33. http://dx.doi.org/10.3390/agriengineering6010002
  » http://dx.doi.org/10.3390/agriengineering6010002
• Herr AW, Adak A, Carroll ME, Elango D, Kar S, Li C, et al. 2023. Unoccupied aerial systems imagery for phenotyping in cotton, maize, soybean, and wheat breeding. Crop Science 63: 1722-1749. http://dx.doi.org/10.1002/csc2.21028
  » http://dx.doi.org/10.1002/csc2.21028
• Herrera D, Escudero-Villa P, Cárdenas E, Ortiz M, Varela-Aldás J. 2024. Combining image classification and unmanned aerial vehicles to estimate the state of explorer roses. AgriEngineering 6: 1008-1021. http://dx.doi.org/10.3390/agriengineering6020058
  » http://dx.doi.org/10.3390/agriengineering6020058
• Jia Y, Fu K, Lan H, Wang X, Su Z. 2024. Maize tassel detection with CA-YOLO for UAV images in complex field environments. Computers and Electronics in Agriculture 217: 108562. http://dx.doi.org/10.1016/j.compag.2023.108562
  » http://dx.doi.org/10.1016/j.compag.2023.108562
• Kumar A, Desai SV, Balasubramanian VN, Rajalakshmi P, Guo W, Naik BB, et al. 2021. Efficient maize tassel-detection method using UAV based remote sensing. Remote Sensing Applications: Society and Environment 23: 100549. http://dx.doi.org/10.1016/j.rsase.2021.100549
  » http://dx.doi.org/10.1016/j.rsase.2021.100549
• Lambert RJ, Johnson RR. 1978. Leaf angle, tassel morphology, and the performance of maize hybrids. Crop Science 18: 499-502. http://dx.doi.org/10.2135/cropsci1978.0011183x001800030037x
  » http://dx.doi.org/10.2135/cropsci1978.0011183x001800030037x
• Li Y, Ma R, Zhang R, Cheng Y, Dong C. 2023. A tea buds counting method based on YOLOv5 and Kalman filter tracking algorithm. Plant Phenomics 5: 0030. http://dx.doi.org/10.34133/plantphenomics.0030
  » http://dx.doi.org/10.34133/plantphenomics.0030
• Liu S, Yin D, Feng H, Li Z, Xu X, Shi L, et al. 2022. Estimating maize seedling number with UAV RGB images and advanced image processing methods. Precision Agriculture 23: 1604-1632. http://dx.doi.org/10.1007/s11119-022-09899-y
  » http://dx.doi.org/10.1007/s11119-022-09899-y
• Lu S, Xuan J, Zhang T, Bai X, Tian F, Ortega-Farias S. 2022. Effect of the shadow pixels on evapotranspiration inversion of vineyard: a high-resolution UAV-based and ground-based remote sensing measurements. Remote Sensing 14: 2259. http://dx.doi.org/10.3390/rs14092259
  » http://dx.doi.org/10.3390/rs14092259
• Lu C, Nnadozie E, Camenzind MP, Hu Y, Yu K. 2024. Maize plant detection using UAV-based RGB imaging and YOLOv5. Frontiers in Plant Science 14: 1274813. http://dx.doi.org/10.3389/fpls.2023.1274813
  » http://dx.doi.org/10.3389/fpls.2023.1274813
• Niu S, Nie Z, Li G, Zhu W. 2024. Multi-altitude corn tassel detection and counting based on UAV RGB imagery and deep learning. Drones 8: 198. http://dx.doi.org/10.3390/drones8050198
  » http://dx.doi.org/10.3390/drones8050198
• Pereira Filho IA, Borghi E. 2022. Disponibilidade de Cultivares de Milho para o Mercado de Sementes do Brasil: Safra 2021/2022. Embrapa Milho e Sorgo, Sete Lagoas, MG, Brazil (in Portuguese).
• Rodríguez-Lira D-C, Córdova-Esparza D-M, Álvarez-Alvarado JM, Romero-González J-A, Terven J, Rodríguez-Reséndiz J. 2024. Comparative analysis of YOLO models for bean leaf disease detection in natural environments. AgriEngineering 6: 4585-4603. https://doi.org/10.3390/agriengineering6040262
  » https://doi.org/10.3390/agriengineering6040262
• Shi R, Tong L, Du T, Shukla MK. 2020. Response and modeling of hybrid maize seed vigor to water deficit at different growth stages. Water 12: 3289. http://dx.doi.org/10.3390/w12113289
  » http://dx.doi.org/10.3390/w12113289
• Sirisha U, Praveen SP, Srinivasu PN, Barsocchi P, Bhoi AK. 2023. Statistical analysis of design aspects of various YOLO-based deep learning models for object detection. International Journal of Computational Intelligence Systems 16: 126. http://dx.doi.org/10.1007/s44196-023-00302-w
  » http://dx.doi.org/10.1007/s44196-023-00302-w
• Song CY, Zhang F, Li JS, Xie JY, Yang C, Zhou H, et al. 2023. Detection of maize tassels for UAV remote sensing image with an improved YOLOX Model. Journal of Integrative Agriculture 22: 1671-1683. http://dx.doi.org/10.1016/j.jia.2022.09.021
  » http://dx.doi.org/10.1016/j.jia.2022.09.021
• Wang B, Zhou J, Costa M, Kaeppler SM, Zhang Z. 2023. Plot-level maize early stage stand counting and spacing detection using advanced deep learning algorithms based on UAV imagery. Agronomy 13: 1728. http://dx.doi.org/10.3390/agronomy13071728
  » http://dx.doi.org/10.3390/agronomy13071728
• Yin L, Wang L, Li J, Lu S, Tian J, Yin Z, et al. 2023. YOLOV4_CSPBi: enhanced land target detection model. Land 12: 1813. http://dx.doi.org/10.3390/land12091813
  » http://dx.doi.org/10.3390/land12091813
• Yu X, Yin D, Nie C, Ming B, Xu H, Liu Y, et al. 2022. Maize tassel area dynamic monitoring based on near-ground and UAV RGB images by U-Net model. Computers and Electronics in Agriculture 203: 107477. http://dx.doi.org/10.1016/j.compag.2022.107477
  » http://dx.doi.org/10.1016/j.compag.2022.107477
• Zhang G, Qin D, Yang J, Yan M, Tang H, Bie H, et al. 2022. UAV Low-altitude aerial image stitching based on semantic segmentation and ORB algorithm for urban traffic. Remote Sensing 14: 6013. http://dx.doi.org/10.3390/rs14236013
  » http://dx.doi.org/10.3390/rs14236013
• Zhang X, Zhu D, Wen R. 2023. SwinT-YOLO: detection of densely distributed maize tassels in remote sensing images. Computers and Electronics in Agriculture 210: 107905. http://dx.doi.org/10.1016/j.compag.2023.107905
  » http://dx.doi.org/10.1016/j.compag.2023.107905
• Zhao J, Zhang X, Yan J, Qiu X, Yao X, Tian Y, et al. 2021. A wheat spike detection method in UAV images based on improved YOLOv5. Remote Sensing 13: 3095. http://dx.doi.org/10.3390/rs13163095
  » http://dx.doi.org/10.3390/rs13163095
• Zhao Y, Yang Y, Xu X, Sun C. 2023. Precision detection of crop diseases based on improved YOLOv5 model. Frontiers in Plant Science 13: 1066835. http://dx.doi.org/10.3389/fpls.2022.1066835
  » http://dx.doi.org/10.3389/fpls.2022.1066835

Data availability statement:

Once accepted, the data will be available as requested and/or in a public repository.

Declaration of use of AI Technologies:

The authors declare that no AI technologies were used to create this work.

*Corresponding author

<[e-mail]>

Edited by:

Mário Henrique Murad Leite Andrade

Conflict of interest:

The authors declare that they have no competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.

Publication Dates

• Publication in this collection
  27 June 2025
• Date of issue
  2025

History

• Received
  17 Sept 2024
• Accepted
  26 Jan 2025