Wheat spike and spikelet detection and counting from high-resolution digital imagery using YOLO with Oriented Bounding Boxes

Abstract

In-season estimation of wheat grain yield potential is critical for crop management and advancing breeding efforts. Spike and spikelet counts serve as key indicators directly linked to yield potential, yet their assessment still relies on manual counting which is both labor-intensive and error-prone. High-resolution digital (RGB) imagery combined with deep learning-based object detection methods has substantially advanced automatic wheat spike detection and counting. However, precise spikelet-level phenotyping remains largely underexplored. This study evaluates two recent YOLO variants, YOLOv11 and YOLOv12, for wheat spike and spikelet detection and counting using oriented bounding boxes (OBB), and introduces a new large-scale benchmark dataset comprising 48,521 spike and 60,404 spikelet instances with OBB annotations. For spike detection, the pre-trained YOLOv11 achieved superior accuracy (mAP@0.5 = 95.8%, Pearson r = 0.993) with shorter training and inference times compared to YOLOv12. For spikelet detection, the non-pretrained YOLOv11 demonstrated higher accuracy (mAP@0.5 = 99.0%), while counting performance was comparable across models. These results establish OBB-based YOLO detection as a robust and scalable approach for AI-driven wheat phenotyping.

Supplementary Information

The online version contains supplementary material available at https://doi.org/10.1038/s41598-026-55761-w.

Introduction

The global food demand is increasing, driven by population growth, rising incomes, and shifting dietary preferences¹. As one of the most important cereal crops, wheat contributes 15–20% of global dietary protein and remains fundamental to agricultural sustainability and food security worldwide². With cultivation spanning approximately 218 million hectares worldwide, this critical crop faces an anticipated demand surge to three billion tons by 2050³. However, the global food supply faces mounting challenges due to a reduction in arable land, evolving pest and disease dynamics, disparities in access to agricultural advances, and geopolitical instability⁴, hence demanding creative solutions to boost the quantity and quality of cereal crops. Additionally, rising demand is threatened by climatic instability, with projections indicating a 4–6% decline in global wheat production for every 1 °C increase in air temperature⁵. Given these circumstances, improving wheat production through advanced breeding, precision agriculture, and enhanced field management is essential to sustain global supply⁶. In particular, the deployment of advanced phenotyping tools capable of rapidly and accurately assessing key growth traits and estimating grain yield potential in wheat plays a pivotal role in supporting field management practices, optimizing breeding efforts, and ensuring agricultural productivity and long-term sustainability.

Among the key determinants of wheat productivity, spikes per unit area, grain number per spike, and thousand-grain weight are critical. Wheat grain yield potential is estimated as the product of these three key components⁷. However, compensatory trade-offs among these yield components pose a major breeding challenge, as improving one trait often comes at the expense of another due to limited assimilate supply per grain and intensified intra-spikelet resource competition. This challenge underscores the importance of precise, high-throughput phenotyping of spike and spikelet traits, which enables breeders to quantify trade-offs, identify favorable trait combinations, and guide selection strategies for yield improvement⁸. Beyond breeding applications, accurate estimation of these traits provides valuable indicators for both farmers and agronomists, enabling the assessment of crop growth status, in-season estimation of yield before harvest, and the optimization of management practices⁹.

The prevailing method for enumerating wheat spikes and spikelets involves field observations, followed by sample collection and manual counting¹⁰. Manual counting is characterized by imprecision, high cost, and low efficiency, as it depends on expert observation and is therefore susceptible to subjective errors¹¹. These identified limitations in wheat phenotyping hinder efforts to uncover the genetics of quantitative traits, thereby creating a bottleneck in breeding advancements¹². High-throughput plant phenotyping (HTP) has been recognized as integral to overcoming these bottlenecks in measuring plants’ measurable traits. Image-based HTP is increasingly utilized as a non-destructive, rapid, and efficient method for phenotyping crops, including wheat¹³. Specifically in wheat, automated systems that can consistently and reliably process large volumes of imagery to rapidly and accurately estimate spike¹⁴ and spikelet counts at the field-plot scale present a promising high-throughput application.

Prior studies have implemented automated counting of wheat spikes through various image processing and modeling techniques, which utilize high-resolution digital imagery (i.e., RGB images) collected from ground-based close-range digital cameras or low-altitude aerial imaging systems¹⁵. Automated object counting in digital imagery is typically achieved through two approaches. The first rely on traditional image processing methods, such as color-texture analysis¹⁶ or morphology-based techniques¹⁷, which heavily rely on manually extracted static features, such as color, texture, and shape. These hand-crafted, static features are highly sensitive to variations in illumination, background complexity, and canopy structure, limiting robustness and generalization across diverse field conditions. The second involves deep learning-based methods, which are generally more robust as they automatically learn features from images without the need for hand-crafted feature extraction. In addition, they are better suited to handle complex scenes, scale variation, and noisy agricultural environments¹⁸.

Recent studies have applied deep learning-based techniques for wheat spike and spikelet counting from digital imagery, and the specific strategies used can be broadly categorized into: (1) regression-based methods, (2) segmentation-based methods, and (3) object detection-based methods¹⁸. Regression-based techniques learn directly from image features to estimate total spike or spikelet counts without explicitly detecting individual objects. Although relatively simple and computationally efficient, they lack instance-level representation, are often sensitive to scale and density variations, and perform poorly under noisy field conditions, making them less adaptive in real-world environments¹⁹. Segmentation-based techniques, such as Mask R-CNN²⁰, U-Net²¹, and MobileNetV2²², have been widely used in wheat spike studies. These methods partition images into pixel-wise separations between objects and background as distinct regions. Nevertheless, they require extensive pixel-level annotations for training, yet their performance often degrades when target objects are densely distributed, heavily occluded, or overlapping at boundaries, as commonly observed in wheat field imagery¹⁰. In contrast, object detection-based methods combine object localization and classification into a single process, with accuracy and real-time performance serving as key indicators of their algorithmic effectiveness. Such techniques are generally more adaptive to dense and overlapping targets, including wheat spikes and spikelets at varying growth stages and canopy conditions²³. While prior studies have demonstrated the effectiveness of deep learning for wheat spike detection, limitations remain in handling oriented structures, small-object detection, and field-level variability. As a result, detection accuracy may degrade in scenarios involving overlapping spikes or non-uniform orientations, limiting their applicability in complex field conditions.

In terms of object detection methods, especially those utilizing Convolutional Neural Networks (CNNs), are typically divided into two categories: (1) candidate region-based models (two-stage detectors), which first generate regions of interest (RoIs) that serve as potential bounding boxes; features are then extracted from these candidate boxes, often through an RoI pooling layer, to perform both object classification and bounding box refinement; and (2) classification/regression-based models (one-stage detectors), which eliminate the region proposal step and directly predict object categories and bounding box coordinates from the original image in a single stage²⁴. Earlier studies on wheat spike detection from digital imagery have used both one-stage²⁵ and two-stage^26,27 object detection methods. Two-stage detectors, such as Fast R-CNN²⁸, Faster R-CNN²⁹, Mask R-CNN²⁰, and one-stage detection, such as RetinaNet³⁰, Single Shot MultiBox Detector (SSD)³¹, You Only Look Once (YOLO)³², are widely used for spike detection tasks. More recently, transformer-based, end-to-end detectors (e.g., DEtection TRansformer (DETR))³³ have also been explored in the wheat spike detection task³⁴. Recent transformer-based methods offer robust global reasoning and show promise in general object detection, but their performance on dense, small agricultural targets such as spikelets remains underexplored. Unlike natural scene objects, wheat spikelets exhibit extreme intra-class similarity, dense occlusion, and orientation variability that challenge attention mechanisms designed for larger objects. It is essential to note that one-stage detectors are generally faster and more suitable for real-time applications²⁵. Therefore, while future work should include broader comparisons with transformer-based detectors, YOLO provides a practical baseline for field-oriented deployment, making this benchmark useful for establishing reference performance in wheat spike and spikelet detection²⁴.

YOLO-based object detection and counting have become a widely adopted method due to its high accuracy, real-time performance, and strong community-driven development, which has accelerated its evolution from YOLOv1 to the most recent YOLOv12. Previous studies have predominantly employed YOLOv5³⁵, YOLOv7³⁶, and YOLOv8³⁷ for wheat spike detection, demonstrating their effectiveness in agricultural applications. However, the potential of two recent versions of YOLO, i.e., YOLOv11 and YOLOv12, remains largely unexplored. Most existing work has relied on axis aligned bounding box (AABB), while the use of YOLO-based wheat spike and spikelet detection using oriented bounding box (OBB) has been limited. AABBs cannot represent the true orientation and elongated shape of wheat spikes in dense, small and variable-angled scenes, making them less precise than OBBs, which capture orientation directly and reduce overlap and ambiguity³⁸. Although automated detection and counting of wheat spikes from digital imagery have been previously studied, automated spikelet detection and counting remain largely underexplored³⁵. Spikelet-level analysis provides a highly granular and precise indicator of final grain yield, while also enabling accurate and objective quantification of disease severity, such as Fusarium Head Blight (FHB)³⁹. Nevertheless, spikelet detection presents technical difficulties in computer vision due to small targets arranged densely on the rachis, partial occlusion, and self-similarity among spikelets.

To this end, this study contributes by establishing a high-quality, high-resolution, and large-scale benchmark dataset with OBB annotations for both wheat spike and spikelet detection. The spikelet dataset integrates imagery from both field and controlled conditions, bridging a key gap in spikelet-level phenotyping. Building on this dataset, the study systematically evaluates recent YOLO-based object detection models, namely YOLOv11 and YOLOv12, for wheat spike and spikelet detection and counting. Rather than proposing a new detection architecture, the study focuses on adapting and assessing an OBB-based detection framework for fine-grained wheat phenotyping under realistic imaging conditions. The analysis compares detection and counting performance, and evaluates the impact of transfer learning through pretrained YOLOv11 models. In addition, the feasibility of wheat spikelet detection from field-extracted spike images is investigated in the context of fine-grained detection challenges in dense agricultural scenes.

Beyond its technical scope, this work establishes a reproducible, field-relevant framework that integrates OBB detection into large-scale phenotyping pipelines. By enabling accurate, automated spike and spikelet quantification under diverse field conditions, the approach advances the scalability and generalizability of AI-driven crop monitoring. Furthermore, the publicly available benchmark dataset and comprehensive comparative analysis provide a foundation for future algorithmic innovations, encouraging the development of models that can adapt to varying crop types, growth stages, and environmental conditions. In doing so, this study not only addresses a key bottleneck in high-throughput plant phenotyping but also opens new avenues for integrating advanced computer vision into precision agriculture research and practice.

Data

Wheat spike and spikelet image acquisition

The close-range digital images of wheat canopies used for wheat spike detection were collected from winter wheat breeding trials at South Dakota State University (SDSU), Brookings, South Dakota, USA. The images were collected from field trials during the watery ripe stage (Feekes 10.5) in the 2022 and 2023 growing seasons. The trials included the Elite Yield Trials (ELITE) and Advanced Yield Trials (AYT), comprising 36 and 126 genotypes, with 108 and 252 plots per year, respectively. While the full trials consisted of a larger number of plots, this study utilized a subset of 203 images randomly selected from these trials for annotation and analysis. The images represent multiple winter wheat genotypes evaluated under field conditions. High-resolution digital imagery was captured in 2022 with Cedar CT8 × 2 tablet (16-MP Sony sensor, 4608 × 3456) and in 2023 with Cedar CP3 tablet (12-MP Sony IMX386, 4000 × 3000) to establish the SDSU Rotated Wheat Head Data (SD-RWHD). The close-range wheat canopy images were acquired at two different angles, from the top-view (at a height of ~ 30–50 cm above the canopy) and the side-view (at a lateral distance of ~ 30–50 cm from the canopy) (Fig. 1), resulting in a total of 203 field images for the annotation task, representing different varieties, trials, and years.

Fig. 1.

Example images from the datasets used in this study. The figure shows wheat spike samples from three sources/perspectives: (a) SDSU side-view, (b) SDSU top-view, and (c) the Rotated Global Wheat Head Detection (RGWHD) dataset. The spikelet image are samples from (d) the SDSU field-based wheat spike dataset and (e) SDSU lab-based wheat spike data.

To enhance representativeness beyond South Dakota, in addition to field-collected SD-RWHD images, the publicly available Rotated Global Wheat Head Data (RGWHD)¹⁰ was also incorporated to improve dataset diversity for developing the wheat spike detection model. The RGWHD data consists of 700 field-captured images, each with a resolution of 1024 × 1024 pixels, and contains a total of 25,678 instances of wheat spikes. This dataset is a curated subset of 100 randomly selected images from each of the seven categories within the broader Global Wheat Head Dataset (GWHD), which covers a wide range of pedoclimatic conditions^10,40. The combined SD-RWHD field-collected imagery and the RGWHD spike dataset expose models to diverse spike morphologies and canopy architectures representative of global wheat production (Fig. 2).

Fig. 2.

Per-image wheat spike instance distribution in the SD-RWHD (i.e., RGWHD and SDSU dataset combined) and RGWHD datasets. The histograms illustrate the distribution of wheat spike instances per image for the SD-RWHD (left) and RGWHD (right) datasets, respectively. The red dashed line represents the median, and the orange dashed line represents the mean.

For the spikelet dataset, the side-view field images from SD-RWHD (Fig. 1a) were used to extract individual wheat spike clips, resulting in 2,356 isolated spike images for further spikelet-level annotation (see details in the following subsequent section). In addition to these field-obtained spike clips extracted from side-view SD-RWHD images, an additional 1,653 close-range high-resolution images (each containing only one wheat spike) (Fig. 1e) were captured under indoor lighting conditions with an iPhone 14 Pro to enhance the robustness and model generalizability across different image resolutions. In total, 4,009 individual spike images were compiled in this work for spikelet-level annotations, which we refer to as the SDSU Rotated Wheat Spikelet Data (SD-RWSLD). Each dataset contains a single object class, where all annotated instances correspond to wheat spikes or spikelets, respectively.

Wheat spike dataset establishment

The SD-RWHD images were annotated precisely at the spike level with OBBs and class labels for the target object. An open-source cloud-based Computer Vision Annotation Tool (CVAT)⁴¹ was used for annotation with OBB labels on all visible wheat spikes in the image. The dataset contains a single object class referring to spike object. To ensure annotation consistency, an annotation protocol was established, and each image was manually verified by an expert annotator to correct potential inconsistencies. The protocol specifies that bounding boxes are aligned with the principal axis of each spike, tightly enclosing the visible structure while minimizing background inclusion and ensuring consistent handling of occluded and overlapping instances. OBBs were selected over AABBs based on prior evidence demonstrating that OBBs reduce background contamination by 49.51–62.63% in elongated wheat spike structure¹⁰. Unlike AABBs, which tend to include background pixels and adjacent spike regions, OBBs preserve object orientation, enabling more precise localization and reducing representation ambiguity in dense, overlapping canopy conditions. This is critical for downstream spikelet detection and counting, where inter-spike contamination due to neighboring spikes in each cropped spike images would otherwise compromise the accuracy of spikelet per spike in dense field scenarios.

As creating labels for supervised learning is a tedious and time-consuming; instead of manually annotating each image, the annotation pipeline employed a semi-automatic annotation workflow to guide and assist the human labeler. First, a small subset of images was manually annotated to train a YOLOv11-OBB detector, and then the trained detector was used to generate pseudo-labels for unlabeled images. Subsequently, all pseudo-labels were reviewed and corrected under human supervision. The semi-automatic annotation workflow accelerated the labeling process compared to manually annotating the entire dataset exhaustively. Finally, bounding box labels were exported to the YOLO-OBB specification as a nine-element record: an integer class-id followed by the normalized pixel coordinates of the four clockwise vertices (x1, y1, x2, y2, x3, y3, x4, y4). In total, 22,853 wheat spike instances were annotated from 203 field images, i.e., SD-RWHD. This dataset, when merged with the publicly available RGWHD, resulted in a total of 48,521 annotated spikes across 903 images utilized in this study (Table 1).

Table 1.

Summary of the annotated datasets for wheat spike and spikelet detection.

Type	Dataset	Wheat spike		Wheat spikelet
		No. of images	No. of spikes	No. of images	No. of spikelets
Wheat spike dataset	RGWHD	700	25,678	–	–
	SD-RWHD	203	22,853	–	–
	Total	903	48,521	–	–
Wheat spikelet dataset	SD-RWSLD	–	–	4,009	60,404

RGWHD - Rotated Global Wheat Head Data; SD-RWHD - SDSU Rotated Wheat Head Data; SD-RWSLD - SDSU Rotated Wheat Spikelet Data.

The per-image wheat spike instance distribution statistics of the RGWHD and SD-RWHD datasets, as shown in Fig. 2, reveal two distinct but complementary image sets. The SD-RWHD collection, acquired in situ using handheld cameras, contains a median value of 70 and a mean value of 113 wheat spikes per image. The higher mean value indicates the presence of a few crowded scenes in the SD-RWHD dataset. In contrast, the public RGWHD dataset is sparser, with a median of 32 and a mean of 37 spikes per image. The combination of these datasets therefore exposes the detection models to a balanced range of both dense and lightly populated canopy conditions, offering a robust basis for benchmarking wheat spike detection under real-world field scenarios.

Wheat spikelet dataset establishment

While the combination of SD-RWHD and RGWHD data was used for the spike detection task, the side-view wheat canopy images in SD-RWHD (Fig. 1a) contributed to the preparation of the spikelet dataset, i.e., SD-RWSLD. A two-step process was adopted for spikelet annotation. First, the YOLOv11-OBB model was trained on the wheat spike dataset, and the trained detector was applied to the original side-view field imagery in the SD-RWHD dataset to generate bounding box coordinates for detected spike regions. The images were then clipped using the generated bounding boxes with 20-pixel padding from the original bounding box extent to extract individual spikes, while preserving edge context, suppressing background noise, and removing non-crop artifacts. This resulted in 2,356 oriented spike clips. Second, each isolated spike clip was annotated in the CVAT platform with OBB labels, resulting in a total of 30,838 spikelet labels. Figure 3 illustrates the workflow for isolating individual wheat spikes from field images using detected bounding boxes, which are then used to generate a refined dataset for spikelet-level annotation and analysis.

Fig. 3.

End to end field based wheat spike and spikelet detection pipeline. The process starts with side-view field images where oriented spikes are detected, and each spikes are isolated for subsequent spikelet level phenotyping task.

Further, to enhance dataset robustness and the generalizability of the trained model in later stages, an additional 1,653 close-range images captured under indoor conditions were added, containing 29,566 spikelet labels. In total, the spikelet dataset, i.e., SD-RWSLD, comprises 4,009 individual spike images (2,356 field-derived and 1,653 indoor) with 60,404 spikelet instances precisely labeled with OBB.

Similar to the wheat spike annotation workflow, a semi-automated annotation approach was employed for the spikelet-level annotation task. Initially, a small batch of images was manually annotated to train the YOLOv11-OBB detector. The trained spikelet detector was then used to generate pseudo-labels on unlabeled data, which were subsequently reviewed and corrected under human supervision to ensure consistency and accuracy.

Methodology

This study focuses on automatic detection and counting of wheat spikes and spikelets from field-based digital images using OBB-based object detection techniques. Unlike AABBs, which are limited in handling objects with diverse orientations, OBBs align with the natural orientation of spikes and spikelets, enabling the model to accurately capture their morphology and improve detection performance.

Object detection models

In this work, the YOLO object detection algorithm was employed for object localization and classification tasks³². YOLO process the entire image in a single pass to detect objects and their locations, which diverges from the traditional two-stage detection process. YOLOv11 and YOLOv12 represent recent YOLO developments with native OBB support with optimized angle regression loss, making them suitable for oriented agricultural targets. Both models include spatial attention mechanisms that may improve feature representation. Additionally, a pre-trained YOLOv11 model was employed and assessed to evaluate the strength of transfer learning in agricultural object detection tasks.

YOLOv11

YOLOv11 is an evolution of its predecessor, YOLOv8, both of which were developed by Ultralytics⁴². YOLOv11 maintains the foundational backbone-neck-head architecture of previous versions, such as YOLOv5 and YOLOv8, while introducing several key modifications. A schematic of the YOLOv11 architecture is presented in Fig. S1a; a prominent change is the implementation of the Cross-Stage Partial with Kernel Size 2 (C3k2) block, a variation of the Cross-Stage Partial (CSP) bottleneck, within the backbone. This block employs two depth-wise convolutions in the classification branch to enhance efficiency and speed with minimal impact on performance. Another significant innovation is the introduction of the Cross-Stage Partial with Spatial Attention (C2PSA) block following the Spatial Pyramid Pooling Fusion (SPPF) module. This block functions as a spatial attention mechanism, enabling the model to prioritize relevant image regions and improve its ability to detect small and occluded objects. These architectural enhancements are carried through the network’s neck and head, where the C3k2 and C2PSA blocks facilitate efficient feature aggregation and precise object localization. In addition, YOLOv11 retains the same detection head configuration as YOLOv10⁴³. These architectural advancements primarily aim to enhance YOLOv11’s feature extraction capabilities for more accurate localization while improving its overall efficiency and speed.

YOLOv12

YOLOv12 is a recent evolution in the YOLO series, incorporating several architectural enhancements that improve both detection performance and computational efficiency⁴⁴. As depicted in the supplementary Fig. S1b, YOLOv12’s architecture also consists of three main components, namely backbone, neck, andhead. The backbone employs the Residual Efficient Layer Aggregation Network (R-ELAN), which integrates residual connections and adaptive scaling to enhance feature reuse and maintain stable gradient propagation with minimal computational cost. This network also uses 7 × 7 large-kernel separable convolutions to expand the receptive field, improving the detection of small and medium-sized objects. The neck network introduces an Area Attention (A2) mechanism that integrates multi-scale features through advanced spatial information processing. This module uses FlashAttention to reduce memory bottlenecks, thereby enhancing contextual modeling while maintaining low latency. This can improve detection accuracy across various object scales. The detection head adopts a decoupled architecture, separating classification and localization tasks to resolve inherent conflicts between them. It replaces traditional convolutions with a two-layer depth-wise separable convolution in the classification branch, thereby reducing model parameters and increasing computational efficiency. These architectural features provide a foundation for capturing diverse features and contextual information, making YOLOv12 a suitable base for developing accurate and efficient detection models.

Model configuration and hyperparameters

Effective hyperparameter selection is crucial in deep learning, as it directly impacts convergence, generalization, and computational efficiency. In this study, three distinct YOLO variants were evaluated to identify a suitable configuration for the wheat spike and spikelet detection task: (1) a pre-trained YOLOv11 model (denoted as PT-YOLOv11), leveraging weights pretrained on the DOTAv1 dataset⁴⁵, (2) a non-pretrained YOLOv11 model (denoted as YOLOv11), and (3) a non-pretrained YOLOv12 model (denoted as YOLOv12) trained from scratch. As no pre-trained weights were publicly available for the YOLOv12 OBB model while conducting this study, it was trained from scratch.

In the YOLO series, the different variants (n, s, m, l, x) represent a scale of model size, with the extra-large (x) variant having the highest network depth and the most parameters. This increase in model size typically results in a higher mean Average Precision (mAP), the primary metric for evaluating object detection performance. Given the challenging nature of detecting small and fine-grained targets, such as wheat spikes and spikelets, the extra-large (x) variant of both models was employed to leverage their superior capacity for feature learning and detection accuracy. The increased network depth and parameter count are known to improve detection performance at the expense of computational speed and resource requirements. For the PT-YOLOv11 variant, its backbone network weights were initialized from a pre-trained model checkpoint with the goal of improving convergence and generalization. A partial layer freezing strategy was used during fine-tuning, specifically freezing layers 0, 1, and 23 of the backbone. This preserved robust, low-level transferable features while allowing the model to adapt to the domain-specific agricultural dataset.

To ensure a fair comparison, all models were trained and evaluated under identical conditions using the same dataset, data splits, and training protocol. The dataset was randomly partitioned into training, validation, and held-out test sets, enabling unbiased performance evaluation. This setup ensured that performance differences are attributable to model characteristics rather than variations in training conditions. The input image size was set to 1024 × 1024 pixels for spike detection and 512 × 512 pixels for spikelet detection, due to relatively smaller dimensions of the input images. Each model was trained for 2,000 epochs, providing sufficient headroom for convergence, with a mini-batch size of 16. Stochastic gradient descent (SGD) was used as the optimizer with an initial learning rate of 0.01, a momentum of 0.937, and a weight decay of 0.0005. To prevent early training instability, the model training began with a three-epoch linear warm-up phase (warm-up momentum = 0.8). Consistent hyperparameters were applied during initialization to ensure fair and unbiased comparison. Both YOLOv11 and YOLOv12 integrated advanced built-in tools for hyperparameter tuning and automated learning rate scheduling, which further refined the training process. For computational efficiency, automatic mixed precision (AMP) was enabled, along with distributed data parallelism on a dual-GPU setup, which reduced memory consumption and accelerated the model training process.

To strengthen model generalization under heterogeneous field conditions, the default data augmentation pipeline implemented in YOLO was employed. This included a RandAugment approach with hue–saturation–value jitter (Δh = 0.015, Δs = 0.7, Δv = 0.4), horizontal flipping (p = 0.5), ± 10% translation, ± 50% isotropic scaling, and random erasing (p = 0.4). These augmentations were relevant for aerial and ground based agricultural imagery, where illumination, viewpoint, and object scale vary substantially across acquisitions. In addition, mosaic augmentation was retained as it has been shown to improve small-object detection by exposing the model to multi-scale contextual compositions within a single training sample. Model checkpoints were saved after every epoch to document the entire learning progression. Although early stopping was omitted to observe the full learning dynamics, the best-performing model weights evaluated on the validation set were saved alongside the final checkpoint to ensure the reproducibility of both the peak and converged model states.

Model assessment

To evaluate model performance, both object detection and counting metrics were assessed. The object detection metrics assessed the model’s ability to accurately localize and classify targets, whereas the counting metrics evaluate its performance in predicting the total number of objects in an image.

Evaluation metrics for object detection models

The constructed image dataset, along with the corresponding OBB labels, was randomly partitioned into training, validation, and testing sets at a 7:2:1 ratio based on the number of images. The number of wheat spikes and spikelets in each category is described in Table 2.

Table 2.

Dataset split and annotation counts for wheat spike and spikelet detection (OBB labels).

Dataset	Split	No. of images	No. of instances
RGWHD + SD-RWHD	Training	632	33,698
	Validation	180	9,904
	Testing	91	4,929
SD-RWSLD	Training	2,806	42,417
	Validation	802	11,974
	Testing	401	6,013

The dataset was randomly partitioned into training, validation, and testing sets at a ratio of 7:2:1. “Instances” denotes the number of annotated objects.

In YOLO, after the model generates initial bounding box prediction and confidence scores, a non-maximum suppression (NMS) algorithm is applied. This process filters out redundant detections by iteratively selecting the bounding box with the highest confidence score and suppressing all other boxes that overlap with it by more than a predefined intersection over union (IoU) threshold. NMS is applied to ensure that each object is counted only once and that the most accurate localization is retained for subsequent evaluation using metrics such as mAP. Four evaluation metrics were used to assess the object detection model’s detection and classification performance: precision, recall, mAP at an IoU threshold of 0.5 (mAP@0.5) and mAP averaged over IoU threshold from 0.5 to 0.95 (mAP@[0.5:0.95]. mAP@[0.5:0.95] provides a stricter assessment by averaging mAP across multiple IoU thresholds from 50% to 95% at 5% increments. This stringent criterion requires more precise bounding box localization.

For the object detection task, precision measures the ratio of correctly predicted positive instances (true positives, TP) to the total number of positive predictions (true positives plus false positives, TP + FP). Likewise, recall measures the ratio of correctly predicted positive instances (TP) to the total number of actual positive instances (TP + FN). Moreover, mAP is a comprehensive metric that evaluates both detection accuracy and localization quality by integrating precision–recall performance across IoU thresholds. For this study, which involves single-class detection, mAP corresponds to the Average Precision (AP) computed across IoU thresholds.

1

2

3

where  Tis the number of IoU thresholds (from 0.5 to 0.95 at 0.05 intervals), and is the AP computed at a specific IoU threshold. In this study, each dataset contains a single object class; therefore, mAP corresponds to the AP averaged across IoU thresholds rather than across multiple classes. A higher mAP indicates better performance in accurately detecting and localizing objects.

In addition to detection performance, the computational efficiency of each model was also evaluated by analyzing its complexity and operational speed. Key metrics included the number of convolutional layers, total parameters (i.e., total trainable weights and biases), and Giga Floating-point Operations Per Second (GFLOPs). The time required for each stage in the detection pipeline, such as preprocessing, inference, and postprocessing, was also assessed. These metrics are vital for real-time systems as they directly influence the practical utility of these models in automated spike and spikelet detection tasks.

Wheat spike or spikelet counting performance evaluation metrics

To evaluate the model performance in counting spikes and spikelets targets, the YOLO-detected counts were compared against ground-truth counts, and accuracy was quantified using four metrics: (a) root mean square error (RMSE), which measures the square root of the average squared difference between predicted and true counts, giving more weight to larger errors. (b) relative root mean square error (RRMSE) is a measure that normalizes the RMSE by the mean of the observed values, providing a scale-independent measure of error. (c) mean absolute error (MAE), which measures the average of absolute differences between predicted and true values, making it less sensitive to outliers, and (d) Pearson correlation coefficient (r), which measures the linear correlation between true and predicted counts, with a value near 1.0 indicating a strong positive linear relationship, and values near 0 indicating little to no relationship.

Hardware and software configuration

The experiments conducted in this study used a software environment based on Rocky Linux 9.1, with Python 3.12.9 and employed PyTorch 2.2.2 as the deep learning framework to ensure full compatibility with CUDA 12.1. The specific hardware and software configurations are summarized in Table 3. The system, featuring an Intel Xeon Gold 6342 CPU and two NVIDIA A100 GPUs (each with 80 GB of memory), provided the computational power and memory required to handle high-resolution wheat spike imagery and complex detection tasks. To ensure consistency and comparability, all model experiments (training, validation, and testing) were conducted on the high-performance computing node with identical specifications.

Table 3.

Shows the hardware and software specifications of the computing environment.

System environment	Version
Operating system	Rocky Linux 9.1 (Blue Onyx)
CPU	Intel(R) Xeon(R) Gold 6342 CPU@2.80 GHz
GPU	NVIDIA A100 80GB PCIe
Python	3.12.9
Pytorch	2.2.2 + cu121
CUDA	12.1

This robust setup formed the basis for rapid model convergence, efficient memory management, and stable performance for object detection and bounding box regression for densely occurring wheat spikes and spikelets. This consistent setup was used to train and infer across all three model variants: a non-pretrained YOLOv11 and YOLOv12 model trained from scratch, and a PT-YOLOv11 initialized with pre-trained weights.

Results

Wheat spikes detection and counting results

The performance of three YOLO-based object detection models, PT-YOLOv11, YOLOv11, and YOLOv12, was evaluated on the wheat spike testing subset using both object detection and counting measures. The wheat spike detection results, presented in Table 4, show a clear ranking in model performance. The PT-YOLOv11 model demonstrated marginally higher detection accuracy, achieving the highest mAP@0.5 score of 0.958 and an mAP@[0.5:0.95] score of 0.704 on the testing dataset split. Its high precision (0.936) and recall (0.902) also highlight its capability of detecting wheat spikes with high accuracy. For comparison, the YOLOv11 model achieved an mAP@0.5 of 0.936 and a mAP@[0.5:0.95] of 0.693. The YOLOv12 model, despite its architectural improvements, performed similarly to YOLOv11, with a mAP@0.5 of 0.941 and a mAP@[0.5:0.95] of 0.689. Interestingly, YOLOv12’s mAP@0.5 was slightly greater than that of the non-pretrained YOLOv11, but its mAP@[0.5:0.95] was marginally lower at 0.689.

Table 4.

Comparative performance of YOLO models for the wheat spike detection task.

Models	Metrics	Precision	Recall	mAP@0.5	mAP@[0.5:0.95]
PT-YOLOv11	Validation	0.926	0.908	0.956	0.694
	Testing	0.936	0.902	0.958	0.704
YOLOv11	Validation	0.933	0.866	0.938	0.681
	Testing	0.930	0.876	0.936	0.693
YOLOv12	Validation	0.921	0.889	0.941	0.679
	Testing	0.930	0.881	0.941	0.689

In addition to evaluating the YOLO models’ performance using object detection metrics, further assessment was performed by comparing each model’s predicted spike counts with ground-truth counts on a per-image basis in the testing dataset split, as shown in Table 5. This evaluation offers an enhanced and rigorous measure of model reliability for phenotyping applications that require precise spike enumeration. The PT-YOLOv11 model achieved the best counting performance, with the highest Pearson correlation (r) of 0.993 and the lowest error metrics: an RRMSE of 10.31%, an MAE of 3.978, and an RMSE of 5.586. The YOLOv11 model showed slightly lower counting performance, with a Pearson correlation (r) of 0.988, an RRMSE of 13.72%, an MAE of 4.538, and an RMSE of 7.431. In comparison, YOLOv12 exhibited reduced counting performance, with the highest error metrics: an RRMSE of 16.06%, an MAE of 5.440, and an RMSE of 8.698.

Table 5.

Comparative performance of YOLO models for the wheat spike counting task.

Model	Pearson r	MAE	RMSE	RRMSE (%)
PT-YOLOv11	0.993	3.978	5.586	10.31
YOLOv11	0.988	4.538	7.431	13.72
YOLOv12	0.989	5.440	8.698	16.06

Note: All metrics are computed on the testing dataset against ground-truth bounding box counts.

In Fig. 4, scatter plots compare predicted and ground-truth per-image spike counts for PT-YOLOv11, YOLOv11, and YOLOv12. Deviation from the 1:1 line increases at higher spike densities, indicating reduced counting accuracy in crowded scenes. The images with higher number of bounding box instances are primarily from SD-RWHD top-view data, which features higher number of spikes per image frame. This results in stronger occlusion and greater capture height, leading to lower effective image resolution, which likely contributes to the observed dispersion. In the images with lower spike density per image, the majority of points fall tightly along the identity (1:1) line.

Fig. 4.

Comparison of predicted versus actual wheat spike counts across test images for three detection models: YOLOv11 (pre-trained), YOLOv11 (non-pretrained), and YOLOv12 (non-pretrained). Each plot illustrates the agreement between model-predicted counts and ground-truth counts.

The representative detection result on the testing dataset is presented in Fig. 5. The best-performing model, i.e., PT-YOLOv11, was used for inference.

Fig. 5.

Representative detection results of the pre-trained PT-YOLOv11 model on the test dataset. The figure illustrates the model’s performance in localizing wheat spikes from both top-view and side-view image perspectives.

In addition to OBB-based prediction overlays, Eigen-CAM activation-map visualizations were generated for representative wheat spike detection cases to provide qualitative insight into model behavior (Supplementary figure S4). Eigen-CAM is commonly used as a post hoc explainability method for CNN- and YOLO-based object detection because it visualizes the principal components of learned convolutional feature representations without requiring model retraining, architectural modification, or gradient backpropagation^46,47. Its gradient-free nature makes it suitable for object detection models, where outputs include bounding-box localization, objectness, and class-confidence components rather than a single class score, and where gradient-based explanations may be less practical or may produce distorted visualizations under uncertain or incorrect predictions^47,48.

Overall, the Eigen-CAM maps showed stronger activation responses around several visible wheat spike regions, suggesting that the detector emphasized biologically meaningful spike-like structures during prediction. In clearer scenes, high-activation regions were more localized and showed stronger spatial agreement with OBB-predicted spikes. In denser or partially occluded scenes, the activation patterns became more diffuse, with some responses also occurring around leaves, stems, high-contrast background areas, and image boundaries. These results indicate that the model generally attends to relevant spike-bearing regions, while also relying on contextual canopy features in visually challenging cases. Therefore, the heatmaps are presented as qualitative model-introspection tools that complement the OBB overlays, rather than as independent detection-density maps or quantitative counting evidence.

Wheat spikelet detection and counting

The performance of three YOLO-based object detection models, PT-YOLOv11, YOLOv11, and YOLOv12, was evaluated using the wheat spikelet test split by employing both object detection and counting measures. Table 6 summarizes the detection results for wheat spikelets. The non-pretrained YOLOv11 model demonstrated strong performance on the test dataset, achieving a precision of 0.972, a recall of 0.967, a mAP@0.5 of 0.990, and a mAP@[0.5:0.95] of 0.830. This performance was closely matched by the PT-YOLOv11 model, which recorded a precision of 0.970, a recall of 0.972, a mAP@0.5 of 0.990, and a mAP@[0.5:0.95] of 0.828. The YOLOv12 model achieved a precision of 0.967, a recall of 0.973, a mAP@0.5 of 0.989, and a mAP@[0.5:0.95] of 0.825 on the same test dataset split. Although YOLOv12 achieved marginally higher recall, both YOLOv11 variants showed slightly superior overall detection accuracy, as indicated by higher mAP@0.5 and mAP@[0.5:0.95]. Hence, while PT-YOLOv11 and YOLOv11 achieved similar mAP@0.5, YOLOv12 performed slightly lower in terms of this object detection metric.

Table 6.

Comparative performance of YOLO models for the wheat spikelet detection task.

Models	Metrics	Precision	Recall	mAP@0.5	mAP@[0.5:0.95]
PT-YOLOv11	Validation	0.970	0.970	0.990	0.832
	Testing	0.970	0.972	0.990	0.828
YOLOv11	Validation	0.971	0.968	0.990	0.835
	Testing	0.972	0.967	0.990	0.830
YOLOv12	Validation	0.968	0.969	0.989	0.827
	Testing	0.967	0.973	0.989	0.825

In addition to evaluating each model’s performance using the object detection metrics, further assessment was performed by comparing each model’s predicted spikelet counts with ground-truth counts on a per-image basis in the testing dataset. All three models, PT-YOLOv11, YOLOv11, and YOLOv12, achieved comparable counting performance, with an RRMSE value close to 8%. Among them, YOLOv12 showed slightly lower accuracy, with an RRMSE of 8.03%, an MAE of 0.796, and an RMSE of 1.204. Overall, the counting performance of all models was consistent with their detection-based evaluation metrics.

Figure 6 illustrates the counting performance of the three models on the wheat spikelet counting task. As shown in Table 7, all YOLO models achieved high agreement with ground-truth counts, with Pearson’s r ≈ 0.93. The relative RMSE remained close to 8% across models, indicating consistent accuracy. Among them, YOLOv12 yielded slightly higher error values (MAE = 0.796, RMSE = 1.204, RRMSE = 8.03%) compared to PT-YOLOv11 and YOLOv11, while PT-YOLOv11 yielded marginally higher counting accuracy. In the scatterplot, a tendency toward slight overestimation is also evident, as reflected in the scatter points positioned above the identity (1:1) line.

Fig. 6.

Comparison of predicted versus actual wheat spikelet counts across testing images for three detection models: PT-YOLOv11, YOLOv11, and YOLOv12 (left to right). Each plot illustrates the agreement between model-predicted counts and ground-truth counts.

Table 7.

Comparative performance of YOLO models for the wheat spikelet counting task.

Model	Pearson r	MAE	RMSE	RRMSE (%)
PT-YOLOv11	0.928	0.736	1.181	7.87
YOLOv11	0.928	0.751	1.185	7.90
YOLOv12	0.930	0.796	1.204	8.03

Note: All metrics are computed on the testing dataset against ground-truth bounding box counts.

The representative detection result on the test dataset of SD-RWSLD is presented in Fig. 7. The best-performing model, as evaluated by the counting metrics with lowest RRMSE, i.e., PT-YOLOv11, was used for inference.

Fig. 7.

Representative detection results of the pre-trained PT-YOLOv11 model on the test dataset split. The figure demonstrates the model’s performance in localizing wheat spikelets on field-extracted images.

Comparison of model computational cost

This section details the computational characteristics of the YOLO models for wheat spike and spikelet detection. The results, summarized in Tables 8 and 9, include key indicators such as the model parameter size, GFLOPs, training time, and inference speed, which are essential for assessing the trade-offs between model expressiveness and computational overhead. Additionally, from a real-time agricultural deployment perspective, computational efficiency is a decisive factor in selecting models for practical use. For comparison, all analyses were performed in the same computing environment, as shown in Table 3.

Table 8.

Comparison of model parameters for both spike and spikelet detection tasks.

Model	Layers	Parameter size (M)	GFLOPs
PT-YOLOv11	199	58.736	202.700
YOLOv11	199	58.736	202.700
YOLOv12	693	63.232	209.400

Table 9.

Comparison of model training and testing time in the wheat spike detection task.

Model	Model size (MB)	Training time (h)	Train speed (ms/img)*	Inference (Min)	Inference speed (ms/img)*
PT-YOLOv11	113	8.742	12.800	0.106	14.033
YOLOv11	113	9.151	12.400	0.108	14.001
YOLOv12	120	10.917	15.500	0.398	22.770

Note: *Train/Inference (ms/img) = preprocess + inference + post-process time reported per image.

Computational cost analysis for wheat spike detection

As shown in Tables 8 and 9, YOLOv12 exhibits the highest architectural complexity, with 693 layers, 63.232 million parameters, and 209.4 GFLOPs, which accordingly leads to the slowest inference, with 22.77 ms/img and a total inference time of 0.398 min. In contrast, PT-YOLOv11 is more efficient, achieving the fastest training time (8.742 h), total inference time (0.106 min), and faster inference (14.033 ms/img) on the testing dataset, while using fewer parameters (58.736 million) and incurring a lower computational cost (202.7 GFLOPs). A comparison of the two YOLOv11 variants further underscores the benefits of pretraining: although the non-pretrained YOLOv11 processed images slightly faster during training (12.4 ms/img), the pre-trained PT-YOLOv11 completed training in less time overall (8.742 h vs. 9.151 h), reflecting a faster convergence rate.

Computational cost analysis for wheat spikelet detection

Tables 8 and 10 present the computational metrics for PT-YOLOv11, YOLOv11, and YOLOv12 in the context of the wheat spikelet detection task. Notably, this task also demonstrated a computational profile similar to that of spike detection task. The YOLOv12 model exhibits the highest complexity (209.4 GFLOPs) and the slowest inference performance (5.827 ms/img; 0.4096 min), with a notably longer training time of 28.089 h. This reflects the added challenge of detecting smaller and more densely packed spikelet targets. When comparing the two YOLOv11 variants, the PT-YOLOv11 model exhibits substantially lower training time (22.545 h) with an inference time of 0.356 min, achieving an inference speed of 5.212 milliseconds per image. For this specific task, the non-pretrained YOLOv11 model completed training in 22.814 h, with the fastest inference time of 0.323 min. Both YOLOv11 models have the same number of layers (199), parameter size (58.736 million), GFLOPs (202.700), and model size (112 MB).

Table 10.

Comparison of model training and testing time in the wheat spikelet detection task.

Model	Model size (MB)	Training time (h)	Train speed (ms/img)*	Inference (Min)	Inference speed (ms/img)*
PT-YOLOv11	112	22.545	5.900	0.356	5.212
YOLOv11	112	22.814	6.200	0.323	4.602
YOLOv12	119	28.089	8.100	0.409	5.827

Note: *Train/Inference (ms/img) = preprocess + inference + post-process time reported per image.

Discussion

Comparison of YOLO models in wheat spike detection and counting

This study introduces a unified OBB-based phenotyping framework for both wheat spikes and spikelets, supported by a new large-scale dataset comprising 48,521 spikes and 60,404 spikelets. Building on YOLOv11 and YOLOv12 as strong baselines, this study contributes three methodological insights relevant to object detection in plant phenotyping: (1) detection mAP does not necessarily predict counting accuracy in dense wheat imagery, (2) transfer learning substantially improves performance across domain shifts, and (3) spikelet detection emerges as a separate fine-grained task requiring tailored modeling strategies. These findings advance the understanding of detection reliability and evaluation limits in dense, fine-grained plant phenotyping tasks.

When comparing the object detection metrics of the non-pretrained YOLOv11 and YOLOv12 models, YOLOv12 demonstrated a marginal advantage in object detection accuracy, achieving an mAP@0.5 of 0.941 compared with YOLOv11’s mAP@0.5 of 0.936 (Table 4). This outcome may be attributed to YOLOv12’s deeper architecture, characterized by its unique Residual Efficient Layer Aggregation Network (R-ELAN) and Area Attention (A2) mechanism, which may have provided a slight advantage in learning and localizing larger, more visually distinct spike targets within complex canopy backgrounds^44,49. The A2 mechanism, which utilizes FlashAttention for efficient multi-scale feature integration, likely contributed to a more robust semantic understanding beneficial for detecting wheat spikes, which are relatively larger targets. This finding aligns with the proposed framework in the YOLOv12 paper⁴⁴, which reports improved detection performance due to its architectural enhancements. This study, however, did not compare the performance of YOLOv12 against other earlier versions of YOLO, such as YOLOv8 or YOLOv5. Hence, future work should compare the performance of YOLOv12 with earlier models in the YOLO family, such as YOLOv8 or YOLOv5, specifically for the task of wheat spike detection.

When evaluating the counting performance of YOLOv11 and YOLOv12 models, a notable divergence was observed between object detection metrics, such as mAP, and counting metrics, such as the Pearson correlation coefficient (r) and RMSE. While YOLOv12 achieved slightly higher mAP@0.5 (0.941) than YOLOv11 (0.936) in the spike detection task (Table 4), YOLOv11 showed better counting accuracy with relatively lower detection error, with MAE of 4.538, RMSE of 7.431, and RRMSE of 13.72% (Table 5). This discrepancy could arise because object detection metrics, such as mAP, primarily prioritize the precise localization and classification of individual targets, requiring a high IoU with respect to ground-truth bounding boxes. Conversely, counting metrics, such as RRMSE, evaluate the model’s ability to accurately estimate the total number of instances within an image, irrespective of the precise bounding box fit. As a result, a model may excel in precisely localizing objects but still produce slight over- or under-counts, whereas another model with slightly lower localization accuracy might better approximate the true counts.

To illustrate the discrepancy between the object detection metric and counting accuracy, a representative example is shown in Fig. 8. In this case, object detection accuracy is quantified using the IoU metric, while counting accuracy is evaluated using the RRMSE. Interestingly, the results highlight that high localization accuracy does not necessarily translate into superior counting performance. In this specific example, YOLOv12 achieved the highest IoU (0.705), indicating a more precise bounding box alignment; however, it exhibited the weakest counting accuracy, with an RRMSE of 7.2%. In contrast, YOLOv11 and PT-YOLOv11 produced identical counting errors (RRMSE = 3.64%) despite achieving slightly lower IoUs (0.693 and 0.674, respectively). Other studies on YOLO-based fruit detection have also reported a divergence between counting performance and localization accuracy⁵⁰. This suggests that models may differ in their ability to balance detection precision with object counting, and that counting metrics should be considered in conjunction with detection accuracy for agricultural applications where enumeration is critical. Complementary to the quantitative analysis in Fig. 8, qualitative error analysis is provided in the supplementary figure S4 to illustrate representative failure cases across models. The examples highlight common sources of false detections and missed detections, including partial occlusion, interference from leaf and stem structures, variations in illumination, and the presence of small, densely packed spikes. Additional challenges such as motion blur and low-contrast regions further contribute to degraded detection performance, particularly in images with IoU values below 60%.

Fig. 8.

Example illustrating the discrepancy between object detection and counting accuracy. (A) Input image; (B–D) oriented bounding-box detections with ground truth (in blue) overlayed on predictions (in red) from PT-YOLOv11, YOLOv11, and YOLOv12 models, respectively.

Further, to assess the impact of transfer learning on the detection and counting of wheat spikes, a comparison is conducted between PT-YOLOv11 and YOLOv11. In the spike detection task, the PT-YOLOv11 model achieved the highest mAP@0.5 (0.958) and mAP@[0.5:0.95] (0.704) scores (Table 4). Additionally, the wheat spike counting metrics showed that PT-YOLOv11 slightly outperformed YOLOv11, with the highest Pearson correlation (r) (0.993) and the lowest RRMSE (10.31%) (Table 5). This underscores that pretraining not only improves object localization but also enhances quantitative trait estimation, as evident in the spike counting results - particularly in complex field environments where robust feature representations are crucial. Although the DOTAv1 dataset⁴⁵ used for pretraining PT-YOLOv11 originates from aerial remote sensing imagery, its large-scale, orientation-rich annotations could have provided transferable low- and mid-level features that benefit small and oriented object detection in wheat spike data. When there is no severe domain mismatch between the pretraining data and the target domain, transfer learning provides accelerated convergence⁵¹. This result highlights the advantage of utilizing pre-trained weights from a large-scale external dataset, even in cross-domain scenarios.

The learning curves in Fig. 9 highlight the benefit of pretraining, as PT-YOLOv11 converged more rapidly and stabilized sooner than the non-pretrained models. Transfer learning in this setting improved data efficiency by providing strong initial feature representations⁵², which is especially valuable in agricultural domains where assembling large annotated datasets is resource-intensive. Moreover, YOLOv12 exhibited slightly faster convergence than YOLOv11 in the early epochs, likely due to architectural refinements such as R-ELAN and A2 that enhance feature aggregation and attention. This study did not evaluate a pre-trained model based on YOLOv12 because such a model was unavailable at the time of this study. Future work could investigate the potential of pre-trained YOLOv12 to compare its performance with that of the pre-trained YOLOv11 model presented in this paper.

Fig. 9.

Learning curves of YOLO variants (PT-YOLOv11, YOLOv11, YOLOv12) evaluated on wheat spike data, showing mAP@0.5 (left) and mAP@0.5:0.95 (right) across the first 250 training epochs.

Comparison of YOLO models in wheat spikelet detection and counting

In this section, the comparative evaluation of PT-YOLOv11, YOLOv11, and YOLOv12 models is assessed for wheat spikelet detection and counting tasks. When assessing the performance of YOLOv11 and YOLOv12 for spikelet detection, YOLOv11 achieved a mAP@0.5 of 0.990, marginally outperforming YOLOv12’s 0.989 (Table 6). The superiority of YOLOv11 differs from the observation in the spike detection task. Such variation may be attributed to architectural differences between YOLOv11 and YOLOv12, which may have impacted the handling of relatively smaller dimensions of spike-level image chips extracted from larger high-resolution field images. Each individual spike is extracted from a larger field image scene, which requires preserving fine-grained spatial features—a task where excessive network depth may be counterproductive⁵³. These findings are consistent with previous studies that have demonstrated the robustness of YOLOv11 compared with YOLOv12 in terms of accuracy in agricultural applications^49,50,54. While YOLOv12’s A2 mechanism efficiently processes large receptive fields to improve semantic understanding, this process may inadvertently average out the subtle, low-level details crucial for identifying small objects such as spikelets. In contrast, YOLOv11’s more compact architecture and shallower detection head likely minimized this feature abstraction, enabling a more direct flow of high-resolution spatial information. Furthermore, its Cross-Stage Partial with Spatial Attention (C2PSA) module is specifically designed to enhance spatial attention in feature maps, thereby improving accuracy for small and overlapping objects⁵⁵. YOLOv11 and YOLOv12 exhibited comparable counting performance, consistent with their detection-based evaluation metrics. Both models achieved a strong correlation with ground-truth counts (Pearson’s r ≈ 0.93), with RRMSE values of 7.90% and 8.03%, respectively.

To assess the impact of transfer learning on the detection and counting of wheat spikelets, PT-YOLOv11 and YOLOv11 were evaluated. Both PT-YOLOv11 and YOLOv11 showed similar results, with a mAP@0.5 of 0.990 for both (Table 6). This suggests that YOLOv11’s inherent architectural design, particularly its C2PSA module, may be highly effective at capturing the necessary fine-grained features for spikelet detection, even without external pretraining. The limited additional benefit from pretraining for spikelets, especially at earlier epochs (Fig. 10), could be attributed to a potential domain mismatch between the source pretraining dataset and the extremely fine-grained, visually homogeneous characteristics of spikelet images. This observation is consistent with earlier studies, which suggest that features learned from broader object categories may not fully capture the subtle distinctions required for specialized targets, thereby leading to less pronounced gains from transfer learning relative to larger and more distinct objects⁵¹. While the non-pretrained YOLOv11 showed a marginal superiority in detection metrics i.e., mAP@[0.5:0.95] (0.830 vs. 0.828), both PT-YOLOv11 and YOLOv11 performed similarly well in terms of counting accuracy (RRMSE 7.9% each) (Table 7).

Fig. 10.

Learning curves of YOLO variants (PT-YOLOv11, YOLOv11, YOLOv12) evaluated on wheat spikelet data, showing mAP@0.5 (left) and mAP@0.5:0.95 (right) across the first 30 training epochs.

Across all the scatter plots in Fig. 6 showing ground truth against model-predicted count, a tendency to overestimation was observed at higher densities, as indicated by the clustering of points above the identity line in the upper range. This overestimation is likely attributable to the trained model misclassifying sterile spikelets as true detections, given the subtle visual distinctions between sterile and fertile spikelets⁸. While OBB provides a simplified geometric approximation rather than exact biological boundaries, it is sufficient for consistent instance-level localization and counting of spikelets under dense and occluded field conditions. A limitation of this study is the classification of both sterile and fertile spikelets into a single class, which leads to an overestimation of spikelet count. Hence, future work should investigate the distinction between sterile and fertile spikelets by incorporating additional morphological traits, such as shape, size, and area, or by utilizing different imaging modalities, such as CT imaging. Wheat spikelets can have up to 12 floret primordia, yet only 2–5 seeds develop, depending on genotype and environment. Further studies are needed to accurately predict the number of seeds per spikelet, which would improve predictions of grain yield potential.

As observed in Fig. 10, YOLOv12 without pretraining exhibited faster convergence during the early training epochs compared with YOLOv11. This accelerated convergence is consistent with the findings reported in the proposed framework in the YOLOv12 paper⁴⁴. The improved training efficiency may be attributed to architectural refinements, such as the incorporation of A2 and the R-ELAN, which facilitate more effective feature aggregation. Transfer learning did not add any advantage for quick convergence at earlier epochs, unlike the observation in the wheat spike detection task. This may be due to a potential domain mismatch between the source pretraining dataset and the extremely fine-grained, visually homogeneous characteristics in spikelet images.

The contrasting benefit of pretraining between the spike and spikelet detection task is primarily explained by object scale and feature transferability. Spike level features share sufficient morphological similarity with pretraining dataset, providing meaningful initialization advantage. Spikelets, however, are extremely fine-grained and visually homogeneous targets with no close analog in the pretraining domain, making randomly initialized weights more adaptable to this specific task. Additionally, the larger SD-RWSLD dataset (60,404 instances) reduces the data-efficiency advantage that pretraining typically offers, allowing the non-pretrained model to learn domain-optimal representations from scratch.

Direct comparison with prior studies is challenging due to differences in datasets and experimental protocols. While the GWHD dataset is commonly used for horizontal bounding box detection or AABBs, there is currently no established benchmark for OBB-based spike and spikelet detection. In addition, the dataset used in this study contains higher object density under field conditions, which further limits direct comparability. The dataset introduced in this work provides a useful reference for OBB-based spike and spikelet detection and may support more consistent comparisons in future research.

Implications of computational cost

Across both spike and spikelet detection tasks (Tables 8, 9, and 10), a trade-off exists between architectural expressiveness and operational speed. Overall, YOLOv11 had a faster training time than YOLOv12 on the same datasets, with the pre-trained PT-YOLOv11 achieving the shortest training time, despite recording similar per-image throughput. This suggests that pretraining provides a better initialization, especially for the spike detection task. Inference benchmarking further indicates lower latency and reduced computational cost for YOLOv11 variants without observable losses in overall detection effectiveness on this dataset. While YOLOv12 achieves comparable accuracy in some cases, this may reflect gains from its attention modules; however, the added complexity and latency make it less attractive for deployment in resource-constrained environments.

Spikelets representing smaller, denser targets increased the training time and inference latency for all models, and YOLOv12 does not yield any additional benefit in these conditions. By contrast, PT-YOLOv11 maintains a superior balance between precision and recall, as well as runtime efficiency, making it the most practical default for real-time, field-scale phenotyping. Given the heterogeneous behavior across YOLO families, careful model selection and training optimization remain essential for task-specific deployments^49,50,54. This study did not explore lightweight variants within the YOLOv11 and YOLOv12 models (such as nano, small, or medium) to determine their computational versus accuracy tradeoff. Future work could benefit from exploring these lightweight models in more effective and practical edge-deployment applications.

Limitations

Despite the advancements presented, several limitations should be acknowledged. The datasets used in this study are constrained to specific acquisition conditions and experimental settings, which may limit the generalizability of the proposed models to different wheat varieties, growth stages, imaging platforms, and environmental conditions. Variations in illumination, canopy structure, occlusion, and sensor characteristics may affect model robustness when applied beyond the training distribution. It is worth noting that the dataset used for wheat spikelet detection and counting was collected primarily from SDSU breeding trials in Brookings, South Dakota, USA. Although it spans 36 and 126 genotypes, collected from 108 and 252 different plots during the 2022 and 2023 growing seasons respectively, its cross-regional and cross-environmental generalizability has not yet been validated. Future work will focus on (i) collecting side-view spikelet images from diverse agroecological zones and germplasm pools, (ii) evaluating domain-adaptation and transfer-learning strategies to extend model applicability across regions and cultivars, and (iii) contributing to community efforts toward an open OBB-annotated benchmark analogous to GWHD.

While the evaluated YOLO-based models demonstrate strong detection and counting performance at the image level, scalability to large-scale field deployments remains unassessed. Processing high-resolution imagery or continuous data streams from UAVs or ground-based systems may introduce computational and memory constraints, particularly for larger model variants. Real-world applicability has not been fully validated under operational conditions. Practical deployment scenarios may involve motion blur, variable sensor angles, partial occlusions, and hardware limitations on edge devices, all of which can affect inference reliability, latency, and consistency. In addition, this study focuses on recent YOLO-based detectors and does not include a comprehensive comparison with alternative detection paradigms. While this targeted evaluation highlights the capabilities of modern YOLO architectures for oriented spike and spikelet detection, broader comparisons would provide a more complete understanding of relative performance across methodologies. Furthermore, although OBB-based detection provides efficient instance-level localization, it represents a simplified geometric approximation; segmentation-based approaches may better capture fine-grained structural details and will be explored in future work. Finally, the absence of a publicly available pre-trained YOLOv12-OBB model at the time of this study precluded the use of transfer learning for this architecture. This limitation may have affected convergence efficiency and performance relative to models initialized with pre-trained weights.

All experiments in this study were conducted in a high-performance computing environment, which may not directly reflect deployment conditions on resource-constrained devices. However, the underlying models have been integrated into a cloud-based platform (https://wheatai.net⁵⁶), which supports both single-image and bulk processing through a browser-based interface. This enables large-scale inference without requiring GPU resources on the user side and demonstrates practical deployment for high-throughput phenotyping workflows. Nevertheless, deployment on edge agricultural platforms introduces additional challenges related to computational capacity, memory, and power constraints, as well as robustness under real-world conditions such as varying illumination, motion blur, occlusion, and sensor variability.

To enhance scalability and practical applicability, future work will evaluate model performance on camera-mounted edge devices, including UAV-based and ground-based platforms, under real-time operational conditions. In this context, lightweight YOLO variants (e.g., nano and small models) will be explored to optimize the trade-off between detection accuracy and computational efficiency. Model optimization techniques, such as pruning, quantization, and efficient architecture design, will further support deployment in resource-constrained environments by reducing inference latency and memory requirements while maintaining acceptable performance. Additionally, this work focuses on image-level detection and counting; future research will extend the framework to row-level and area-based estimation to enhance agronomic interpretability.

Conclusion

This study presents a deep learning–based workflow for detecting wheat spikes and spikelets using close-range, high-resolution digital imagery captured across diverse field conditions. By leveraging field-imaging tablets and smartphones, this study focuses on a practical pipeline encompassing image acquisition, ground-truth data generation, and model development, with the long-term goal of enabling wheat grain yield prediction. It establishes the foundational tools and methodologies necessary to achieve this objective. Recent YOLO variants, i.e., YOLOv11 and YOLOv12, were evaluated, with transfer learning applied to enhance detection accuracy where applicable. OBBs were incorporated to address challenges posed by tilted and overlapping spikes, particularly in dense or irregular canopies. Experimental results demonstrate that these methods can accurately detect and count wheat spikes and spikelets, with model-derived counts showing strong agreement with ground-truth counts. While YOLOv12 achieved higher accuracy for spike detection, YOLOv11 delivered superior counting performance for both spikes and spikelets, underscoring the distinct nature of detection and counting metrics. The inclusion of spikelet-level phenotyping provides a finer-grained assessment of yield components and crop health, offering tangible benefits for breeding programs and in-field decision-making. Overall, the proposed method reduces the time, labor, and subjectivity associated with manual counting, making it a reliable and scalable solution for field phenotyping. Future work will focus on deploying deep learning-based detection and counting methods on Unmanned Aerial Vehicle (UAV) imagery to obtain a detailed field-level assessment. Similarly, lightweight YOLO variants will be explored on camera-mounted edge devices to enable real-time, in-field operation, as well as expand the approach to other cereal crops and environmental conditions.

Supplementary Information

Below is the link to the electronic supplementary material.

Supplementary Material 1 ^{(9.4MB, docx)}

Author contributions

H.G. contributed to data curation, formal analysis, methodology, software, validation, visualization, and writing of the original draft, and reviewed and edited the manuscript. M.M. contributed to conceptualization, data curation, formal analysis, methodology, software, supervision, writing—review and editing, project administration, and funding acquisition. S.T. contributed to data curation and validation. S.K. and D.J.K. contributed to data curation. M.M.B., K.P., Sh.S., M.S., Sa.S., and U.u.R.J. contributed to data curation and validation. S.K.S. contributed to conceptualization, resources, supervision, writing—review and editing, project administration, and funding acquisition.

Funding

This work was primarily supported by the South Dakota Wheat Commission, along with other funding sources including South Dakota State University Agricultural Experiment Station and USDA-NIFA Hatch Project (SD00H829-25), the USDA-NIFA WheatCAP Project (Award #2022-68013-36439), USDA-ARS agreement 59-0206-2-153 (USWBSI), and the National Science Foundation (NSF Award #2316502).

Data availability

The data supporting the findings of this study are openly available at: https://doi.org/10.5281/zenodo.20215489.

Declarations

Competing interests

The authors declare no competing interests.