EVALUATION OF NEXT-GENERATION YOLO DETECTORS FOR COTTON WEED IDENTIFICATION USING A REGION-SPECIFIC DATASET FROM UZBEKISTAN

ABSTRACT

Weed infestation remains a major constraint in cotton production, driving yield losses and increasing the economic and environmental costs of crop management. To address the need for more efficient and eco-friendly weed control strategies, this study evaluates next-generation YOLO object detectors for multi-class weed detection in cotton fields. A region-specific, annotated RGB image dataset was collected in cotton-growing fields of the Nukus District, Uzbekistan across the 2024–2025 agricultural seasons, comprising 1,081 images and 4,411 labeled objects representing cotton and seven dominant weed species. Using this dataset, we trained and compared the YOLOv10 and YOLOv11 model families under identical training and augmentation settings. YOLOv11s achieved the highest overall accuracy, with mAP@0.5 of 0.847 and mAP@[0.5:0.95] of 0.651, while YOLOv10s provided the best accuracy–efficiency trade-off for resource-constrained platforms. These findings offer a baseline for deploying next-generation YOLO detectors in Central Asian cotton weed management.

АННОТАЦИЯ

Засорённость посевов хлопчатника сорными растениями остаётся одним из ключевых факторов, ограничивающих продуктивность культуры, приводя к снижению урожайности и росту экономических и экологических издержек её возделывания. В настоящем исследовании оцениваются модели обнаружения объектов нового поколения семейства YOLO для многоклассового обнаружения сорной растительности на хлопковых полях. Регионально ориентированный аннотированный набор RGB-изображений был собран на хлопкосеющих полях Нукусского района Республики Узбекистан в сельскохозяйственные сезоны 2024–2025 годов и включает 1 081 изображение и 4 411 размеченных объектов, представляющих хлопчатник и семь доминирующих видов сорных растений. На основе данного набора данных были обучены и сопоставлены модели семейств YOLOv10 и YOLOv11 при идентичных условиях обучения и аугментации данных. Модель YOLOv11s продемонстрировала наивысшую суммарную точность, достигнув значений mAP@0.5 = 0.847 и mAP@0.5:0.95 = 0.651, тогда как YOLOv10s обеспечила наилучший баланс между точностью и вычислительной эффективностью на платформах с ограниченными ресурсами. Полученные результаты могут служить базовым ориентиром для развёртывания детекторов семейства YOLO нового поколения в системах управления сорной растительностью в хлопководстве Центральной Азии.

Keywords:  YOLOv10, YOLOv11, weed detection, cotton production, deep learning, computer vision, mAP, precision agriculture, image dataset, Uzbekistan.

Ключевые слова: YOLOv10, YOLOv11, выявление сорняков, производство хлопка, глубокое обучение, компьютерное зрение, mAP, точное сельское хозяйство, набор данных изображений, Узбекистан.

Introduction

Uncontrolled weed growth in cotton fields competes with the crop for water, nutrients and light. It can also promote pest and disease outbreaks. This leads to substantial yield losses and stimulates excessive herbicide use, with negative consequences for soil health and agro-ecosystem biodiversity [1]. Recent system-level analyses emphasise that site-specific weed management and data-driven weed mapping are central to reducing herbicide inputs and environmental impacts in modern cropping systems [2].

Vision-based precision weeding aims to localize weeds at fine spatial resolution and to drive variable-rate or spot spraying, thereby reducing chemical inputs while maintaining control efficacy. Deep convolutional neural networks, especially one-stage detectors in the YOLO family have become the de facto standard for real-time weed detection, thanks to their balance of accuracy and inference speed. Benchmark studies such as YOLOWeeds, which evaluates multiple YOLO variants on cotton weed images from the CottonWeedDet12 dataset, demonstrate that YOLO-based detectors can robustly distinguish multiple weed species across growth stages under field conditions [3]. Beyond cotton, YOLO-based detectors have been successfully used for multi-class weed detection in rice, soybean and tomato fields, often outperforming two-stage detectors in both accuracy and inference speed [2, 4, 5].We identify two main gaps in the current literature. First, weed flora and management practices differ substantially across regions, meaning that detectors trained on foreign datasets may generalize poorly to Central Asian cotton systems. Second, the latest detector generations YOLOv10 and YOLOv11 remain under-explored in agricultural applications compared with YOLOv5–YOLOv9 and have not yet been systematically evaluated for multi-class weed detection in cotton fields using region-specific data [1, 4-6].

The recently introduced YOLOv10 and YOLOv11 families incorporate architectural changes intended to improve the accuracy–efficiency trade-off, including enhanced multi-scale feature aggregation and refined loss designs [7, 8]. Building on our earlier work with YOLOv8 and YOLOv9, this study extends the analysis to these next-generation detectors using the same Nukus District, Uzbekistan dataset and experimental conditions for a fair comparison [9].

The objectives of this study are (1) to utilize a locally relevant, annotated image dataset of cotton and dominant weed species for training and evaluation (2) to benchmark the performance of YOLOv10 and YOLOv11 architectures for multi-class weed detection under realistic field conditions (3) to analyze global, class-wise, and threshold-dependent behavior of representative models from both families and (4) to derive practical recommendations for selecting appropriate detectors for real-time precision weed management systems in Uzbekistan.

Materials and methods

Dataset and image acquisition. RGB images were collected from cotton fields located in Nukus District, Uzbekistan over two seasons to capture phenological diversity and field variability.  A first acquisition campaign produced 225 images in April-May 2024, approximately 15-20 days after planting, targeting early-stage cotton and early weed emergence.  A second campaign in June-August 2025 collected an additional 856 images across multiple cotton growth stages to capture broader morphological variability relevant to detector generalization. Images were captured at camera heights of 50 cm and 100 cm, using near-nadir angles of approximately 80°-90°, reflecting plausible viewpoints for mobile robotic perception systems.  The acquisition device was a Xiaomi 11 Lite 5G NE smartphone producing high-resolution images (6944 × 9280 pixels) stored in JPG format. In contrast to UAV-based weed detection pipelines that operate at several metres above the canopy [4, 5], our low-altitude smartphone acquisition emphasises fine-scale detail and occlusion patterns that are representative of mobile ground robots navigating between cotton rows.

Annotation and data partitioning. All images were annotated in Roboflow using tight bounding boxes around visible plant instances.  The dataset includes eight categories: cotton (Gossypium spp.), Digitaria sanguinalis, Reed, Field bindweed, Xanthium strumarium, Lambsquarters, Abutilon theophrasti, and “Other weeds.”  After initial labeling, annotations were manually reviewed, incorrect labels were corrected, blurred or low-quality samples were removed to improve dataset integrity for training and evaluation. The final curated dataset contains 1,081 images and 4,411 bounding-box annotations (Table 1). Compared with the widely used CottonWeedDet12 benchmark, which comprises 5,648 images and 9,370 bounding boxes across 12 weed species from U.S. cotton systems [3], our dataset is smaller but tailored to the weed flora and management practices of Central Asian cotton fields. It also differs in imaging platform, relying on low-altitude smartphone views rather than UAV or tractor-mounted cameras, which better approximates the perspective of potential ground robots.

Table 1.

 Distribution of labeled instances across target classes.

Images were resized to 640×640 to match standard YOLO training conventions and facilitate fair cross-model comparison. The final partition was manually balanced by season and growth stage to mitigate split bias, resulting in 744 training images, 167 validation images, and 170 test images.

Data augmentation. To improve robustness against variable field conditions (illumination changes, occlusion, camera pose, etc.), data augmentation was applied in two stages: offline (pre-training) and online (during training). All geometric transformations (rotation, translation, scaling, mosaic, copy–paste) were consistently applied to images and bounding boxes.

Offline augmentation (pre-export). Using Roboflow, the base dataset was expanded with rotations at 90° increments, additional random rotations between -13° and +13°, hue shifts in the range -23 to +23, saturation changes between -30 % and +30 %, and brightness enhancement of 0-15 %. These operations increased diversity in color and orientation while preserving object identity.

Online augmentation (training-time). The Ultralytics YOLO training engine performed further on-the-fly transformations, including mosaic augmentation (disabled in the final epochs), horizontal flips with probability 0.5, HSV augmentations (h = 0.015, s = 0.7, v = 0.4), translations up to 0.1 of the image sizes, scaling up to 0.5, random erasing with probability 0.4, copy-paste operations, and RandAugment-style transformations. These online augmentations were applied uniformly across all YOLOv10 (n, s, m, l, x) and YOLOv11 (n, s, m, l) variants to ensure a fair comparison.

Experimental setup and training configuration. All models were trained using the Ultralytics YOLO framework on a workstation equipped with an RTX 4090 (24 GB VRAM), Intel Core i9-12900K CPU, 128 GB DDR5 RAM, and dual NVMe SSD storage. 

Each YOLO model was trained under identical hyperparameters to ensure fair evaluation. The hyperparameters are detailed in Table 2. This protocol follows recent comparative studies where YOLOv9, YOLOv10 and RT-DETR were trained under harmonised settings for weed detection in smart-spraying scenarios, in order to isolate the effect of the detector architecture itself [10].

Table 2.

Hyperparameters for training YOLOv10 and YOLOv11 models

Results and discussion

Table 3 summarizes the global detection performance of the YOLOv10 and YOLOv11 model families for multi-species weed detection in cotton fields. Looking across the YOLOv10 variants, precision remains consistently high, recall is at a comparable level, and mAP@0.5 values indicate reliable detection under field conditions. Among these models, YOLOv10s provides the best overall balance between accuracy and efficiency, with precision around 0.85, recall around 0.76, and mAP@0.5 above 0.83, while maintaining a competitive mAP@[0.5:0.95]. This makes YOLOv10s a reasonable choice when inference speed and resource usage are important constraints, for example on embedded or mobile platforms that cannot support very large detectors.

Table 3.

Evaluation metrics of YOLOv10 and YOLOv11 architectures for weed detection

The YOLOv11 models generally outperform the YOLOv10 family in most global metrics under the same training and augmentation settings. In particular, YOLOv11s achieves the highest overall detection accuracy, with slightly higher precision than YOLOv10s, improved mAP@0.5 and mAP@[0.5:0.95], and comparable recall. These gains suggest that the architectural changes introduced in YOLOv11 such as updated backbone blocks and attention mechanisms translate into more precise localization, especially at stricter IoU thresholds. From a practical point of view, YOLOv11s is therefore better suited for high-performance precision weed management systems where the main objective is to maximize detection quality, while YOLOv10s remains attractive for edge scenarios where computational budgets are limited.

These findings are consistent with our previous comparative study on the same Nukus District dataset, where YOLOv9c surpassed YOLOv8m in overall detection accuracy and mAP@0.5, while YOLOv8m was still identified as a good compromise for energy-efficient hardware [9].

Compared with cotton-focused benchmarks such as CottonWeedDet12, where improved YOLOv11 variants like LMS-YOLO11n achieve modest mAP gains by introducing multi-scale feature fusion modules [3, 11], our results provide a baseline for unmodified YOLOv10 and YOLOv11 architectures in a Central Asian cotton context. This distinction is important: while LMS-YOLO11n optimises the network specifically for a large, public cotton dataset, our experiments isolate the intrinsic strengths of the standard YOLOv10/YOLOv11 designs on a smaller, region-specific dataset that reflects the weed flora and field management of Uzbekistan.

Looking ahead, an obvious next step is to combine these baseline results with the improved YOLOv8/YOLOv9 architectures reported in the literature and to design cotton-specific YOLOv11 variants, followed by field trials on real robotic or UAV platforms to verify performance under fully operational conditions.

Future work will focus on adapting recently proposed YOLOv11-based architectures such as LMS-YOLO11n, HDMS-YOLO and EDGE-MSE-YOLOv11, which introduce multi-scale feature fusion, attention-guided refinement and lightweight perception mechanisms, to our cotton weed scenario [1, 11, 12].

These modules have proven effective in enhancing small-weed and disease detection in rice and mixed-crop fields; transferring and re-tuning them on our dataset may further improve recall for small or early-stage weeds without sacrificing real-time performance.

Another promising direction is to extend our smartphone-based detection system into a full pipeline for site-specific weed management, similar to maize and soybean workflows that combine improved YOLOv11n models with UAV GPS data to generate weed density maps and variable-rate spraying prescriptions [2, 4-6].

In our context, such a pipeline could enable the integration of ground robots or UAVs with local cotton production systems in Uzbekistan, reducing herbicide inputs while maintaining effective weed control.

References:

1. Hua J., He R., Zeng Y., Chen Q. HDMS-YOLO: A multi-scale weed detection model for complex farmland environments // Frontiers in Plant Science. 2025. Vol. 16. Article 1696392. DOI: 10.3389/fpls.2025.1696392.
2. Gómez A., Moreno H., Valero C., Andújar D. Spatio-temporal stability of intelligent modeling for weed detection in tomato fields // Agricultural Systems. 2025. Vol. 228(C).
3. Dang F., Chen D., Lu Y., Li Z. YOLOWeeds: A novel benchmark of YOLO object detectors for multi-class weed detection in cotton production systems // Computers and Electronics in Agriculture. 2023. Vol. 205. Article 107655. DOI: 10.1016/j.compag.2023.107655.
4. Li Y., Guo Z., Sun Y., Chen X., Cao Y. Weed detection algorithms in rice fields based on improved YOLOv10n // Agriculture. 2024. Vol. 14, № 11. Article 2066. DOI: 10.3390/agriculture14112066.
5. Yue Y., Zhao A. Soybean seedling-stage weed detection and distribution mapping based on low-altitude UAV remote sensing and an improved YOLOv11n model // Agronomy. 2025. Vol. 15, № 12. Article 2693. DOI: 10.3390/agronomy15122693.
6. Chen X., Zhang H., Liu X., Guo Z., Zheng W., Cao Y. Research on accurate weed identification and a variable application method in maize fields based on an improved YOLOv11n model // Agriculture. 2025. Vol. 15, № 23. Article 2456. DOI: 10.3390/agriculture15232456.
7. Wang A., Chen H., Liu L., Chen K., Lin Z., Han J., Ding G. YOLOv10: Real-Time End-to-End Object Detection // Advances in Neural Information Processing Systems (NeurIPS 2024): [Электронный ресурс]. 2024. Режим доступа: arXiv:2405.14458.
8. Ultralytics. YOLO11 models: Ultralytics YOLO documentation: [Электронный ресурс]. 2024. Режим доступа: https://docs.ultralytics.com/models/yolo11/
9. Uteuliev N.U., Dauletnazarov J.I. Comparative YOLOv8–YOLOv9 performance in weed recognition for cotton fields // International Journal of Advanced Technology and Natural Sciences. 2025. Vol. 6, № 4.
10. Saltık A.O., Allmendinger A., Stein A. Comparative analysis of YOLOv9, YOLOv10 and RT-DETR for real-time weed detection // Computer Vision – ECCV 2024 Workshops: Milan, Italy, September 29–October 4, 2024: proceedings. Part III. Cham: Springer, 2025. P. 177–193. DOI: 10.1007/978-3-031-91835-3_12.
11. Zhang Y., Xu Y., Hou J., Song Y. LMS-YOLO11n: A Lightweight Multi-Scale Weed Detection Model // International Journal of Advanced Computer Science and Applications (IJACSA). 2025. Vol. 16, № 1. DOI: 10.14569/IJACSA.2025.01601123.
12. Zhang X., Wei L., Yang R. TriPerceptNet: A lightweight multi-scale enhanced YOLOv11 model for accurate rice disease detection in complex field environments // Frontiers in Plant Science. 2025. Vol. 16. Article 1614929. DOI: 10.3389/fpls.2025.1614929.