PaperlayerView exact revision on arXiv

Source-linked AI summary

YOLOv1 to YOLOv10: A comprehensive review of YOLO variants and their application in the agricultural domain

Mujadded Al Rabbani Alif, Muhammad Hussain

arXiv:2406.10139v1cs.CV

TL;DR

Agricultural applications of YOLO variants face domain complexity, regional dataset specificity, and deployment constraints. This survey reviews YOLOv1–YOLOv10 and synthesizes their agricultural uses, finding adaptable, accurate, and real-time support for precision farming, crop monitoring, weed detection, and related tasks, while identifying challenges for broader deployment.

  • Problem

    Agricultural applications involve complex sub-domains, regional dataset limitations, and hardware constraints that complicate broad deployment of YOLO variants.

  • Method

    The survey reviews YOLOv1–YOLOv10, assesses their incremental developments, and synthesizes applications across agricultural tasks.

  • Results

    YOLO-based models show adaptable and consistently high performance across weed detection, crop identification, disease diagnosis, crop monitoring, and real-time precision agriculture.

  • Takeaways & Limitations

    YOLO variants offer real-time and efficient solutions with potential for precision farming, weed management, crop monitoring, and sustainable smart agriculture.

  • Takeaways & Limitations

    Practical deployment remains constrained by resource-limited hardware, regional datasets, and the need to improve generalizability across crops and environmental conditions.

Abstract

from arXiv · show

This survey investigates the transformative potential of various YOLO variants, from YOLOv1 to the state-of-the-art YOLOv10, in the context of agricultural advancements. The primary objective is to elucidate how these cutting-edge object detection models can re-energise and optimize diverse aspects of agriculture, ranging from crop monitoring to livestock management. It aims to achieve key objectives, including the identification of contemporary challenges in agriculture, a detailed assessment of YOLO's incremental advancements, and an exploration of its specific applications in agriculture. This is one of the first surveys to include the latest YOLOv10, offering a fresh perspective on its implications for precision farming and sustainable agricultural practices in the era of Artificial Intelligence and automation. Further, the survey undertakes a critical analysis of YOLO's performance, synthesizes existing research, and projects future trends. By scrutinizing the unique capabilities packed in YOLO variants and their real-world applications, this survey provides valuable insights into the evolving relationship between YOLO variants and agriculture. The findings contribute towards a nuanced understanding of the potential for precision farming and sustainable agricultural practices, marking a significant step forward in the integration of advanced object detection technologies within the agricultural sector.

1 Introduction

This introduction frames YOLO variants as efficient computer-vision tools for precision agriculture and outlines a review of their evolution, applications, performance, and future directions.

  • YOLO and agricultural context: YOLO performs real-time detection by dividing an image into a grid and simultaneously predicting bounding boxes and class probabilities.The review presents this as a major shift from conventional two-stage methodologies.
  • Review significance: The review aims to clarify how YOLO variants may support precision farming and sustainable agricultural practices.The paper organizes this discussion around agricultural challenges, CNN foundations, object detection, YOLO evolution, applications, and future directions.
  • Agricultural needs: Agriculture requires efficient tools for monitoring crops, livestock, and environmental conditions, including crop monitoring, disease detection, yield estimation, and livestock management.YOLO’s real-time capacity, accuracy, and adaptability are presented as relevant to these tasks.
  • Benefits of computer vision: Computer vision supports large-scale precision, resource optimisation, real-time response, and data-driven agricultural decision-making.The review connects these capabilities with monitoring landscapes, allocating water and fertilizers, responding to pests or diseases, and improving farm management.
  • Review scope: The survey traces YOLO’s architectural and algorithmic evolution from YOLOv1 through YOLOv10 and examines applications across agricultural sub-domains.Its stated objectives include assessing performance, synthesizing existing research, and projecting future trends.

2 Convolutional Neural Networks (CNN)

This section introduces CNNs as staged visual-processing architectures whose convolution, pooling, activation, and fully connected components transform image inputs into outputs.

  • CNN architecture: CNNs use convolution, pooling, and activation functions in a staged process to transform inputs into outputs.Their abstract structure includes convolutional blocks and fully connected layers leading to the output.
  • Feature extraction: CNN kernels or filters extract low-level spatial information that subsequent layers use to develop semantic relationships.The number and dimensions of filters are fundamental design choices within convolutional blocks.
  • Convolution operation: Convolution combines weights and spatial inputs through a dot product, adds a bias, and applies a nonlinear activation function.The review identifies weights, inputs, bias, and the activation function as components of this operation.
  • Pooling: Pooling aggregates feature responses by downsampling spatial dimensions, with max pooling given as one available pooling framework.Average pooling and sum pooling are also mentioned.
  • Activation functions: ReLU is widely used in convolutional blocks because its max-operation form is computationally lightweight.The section contrasts ReLU with sigmoid and tanh activation functions.

3 Object Detection

This section contrasts object-detection strategies, emphasizing the computational burden of two-stage methods and the speed and lightweight footprint of single-stage detectors such as YOLO.

  • Detection challenges: Object detectors must handle varying image resolutions, aspect ratios, object sizes, and class imbalance.These factors can produce biased predictions when some classes have insufficient training images.
  • Computational demands: Object detection architectures can require substantial power, memory, and processing time for complex image datasets.The section links this computational complexity to deep internal networks used for feature extraction.
  • Two-stage detectors: Two-stage detectors first propose candidate regions and then classify and localize objects within those regions.Examples include RCNN, Fast R-CNN, Faster R-CNN, and FPN.
  • Two-stage detector development: RCNN’s two-stage process was computationally inefficient, while Fast R-CNN improved speed through ROI pooling.ROI pooling extracts fixed-size feature maps for each region from original feature maps.
  • Single-stage detectors: Single-stage detectors identify objects in one pass without a separate region-proposal step, including SSD, YOLO, RefineDet++, DSSD, and RetinaNet.SSD predicts bounding boxes and class probabilities from multiple convolutional feature maps at different scales.
  • YOLO rationale: Single-stage detectors generally provide faster inference and a lighter footprint than two-stage detectors, making them suitable for resource-constrained environments.YOLO is presented as combining real-time inference capabilities with impressive accuracy and a straightforward architecture.

4 YOLO Architecture Background

YOLO treats object detection as a single-pass regression task, using grid cells, bounding boxes, confidence scores, and non-maximum suppression to identify and localize objects. Across YOLOv1 and YOLOv2, architectural and training changes improved speed, accuracy, recall, scalability, and handling of varied object sizes.

  • YOLO detection principle: YOLO uses one convolutional neural network to predict object locations and bounding boxes directly, framing detection as a regression problem.This single-pass approach distinguishes YOLO from two-stage detection methods.
  • YOLOv1 structure: Each grid cell predicts B bounding boxes with dimensions and confidence scores when an object center falls within that cell.The confidence score represents the probability that an object resides within the designated bounding box.
  • YOLOv1 structure: YOLOv1 applies non-maximum suppression to overlapping predictions, while its loss function accounts for coordinates, dimensions, confidence, and class probabilities.The original architecture used Darknet variants with 24 convolutional layers or nine layers in Fast YOLO.
  • YOLOv1 performance: 63.4% mAP at 45 FPS was achieved by YOLOv1, while Fast YOLO reached 52.7% mAP at 155 FPS.Both variants remained below contemporary state-of-the-art benchmarks, with lower recall and localization errors motivating later improvements.
  • YOLOv2 architecture: YOLOv2 uses Darknet-19, combines pooling with 1 x 1 convolutions, and expands category coverage through ImageNet, COCO, and Word-Tree.The combined datasets broadened detection capabilities to over 9418 object instances.
  • YOLOv2 improvements: YOLOv2 improves detection through data augmentation, higher-resolution classification, Batch Normalization, coordinate prediction, anchor boxes, and K-means clustering.Reported gains include 4% mAP from the higher-resolution classifier, approximately 2% mAP from Batch Normalization, 5% mAP from coordinate prediction, and 7% recall from anchor boxes.

4.3 YOLOv3

YOLOv3 expanded YOLO’s architecture with multi-scale prediction, residual feature extraction, and binary cross-entropy classification while retaining real-time processing. These changes produced a 2.7% AP50 improvement, with YOLOv3 reaching 36.2% AP and 60.6% AP50 at 20 FPS.

  • Architecture: YOLOv3 introduced an expanded architecture that maintained real-time processing while adding objectness scores and four-coordinate predictions for each bounding box.Objectness scores indicate whether an anchor box has the highest overlap with the ground truth.
  • Architecture: Binary cross-entropy replaced SoftMax classification, enabling multiple labels per box, while a 53-layer feature extractor used residual connections.The architecture also incorporated an extensive feature extractor with 53 convolutional layers.
  • Multi-scale detection: 2.7% improvement in AP50 resulted from modified spatial pyramid pooling and three-scale feature maps with prior boxes at multiple resolutions.The feature-map scales were 13×13, 26×26, and 52×52 for a 416×416 input.
  • Performance: YOLOv3 achieved 36.2% AP and 60.6% AP-50 at 20 FPS on MS COCO, establishing Darknet-53 as a benchmark backbone for subsequent YOLO models.The evaluation used the MS COCO dataset after the benchmark transition from PASCAL VOC.
  • YOLOv4 comparison: YOLOv4 later used multiple anchor points for one ground truth, improving positive-sample selection, reducing class imbalance, and enhancing boundary detection accuracy.Its CIoU loss incorporated IoU, maximum IoU, and regularization for localization.

4.5 YOLOv5

YOLOv5 emphasized accessible PyTorch development, feature aggregation, anchor-based prediction, and deployment compatibility. Its reported evaluations showed strong accuracy–speed performance, while YOLOv6 extended the progression with industrial real-time design and separated classification and regression heads.

  • YOLOv5 architecture: YOLOv5 used CSP-based feature extraction, a PAN neck, upsampling, convolutional prediction heads, and anchor-based bounding-box predictions.Its loss combined binary cross-entropy for classification and objectness with CIoU for localization.
  • YOLOv5 training: YOLOv5 combined weighted classification, objectness, and localization losses through loss = λ1 · Lcls + λ2 · Lobj + λ3 · Lloc.The three terms represent binary cross-entropy classification, binary cross-entropy objectness, and CIoU localization losses.
  • YOLOv5 performance: 50.7% AP and 200 FPS were reported for YOLOv5x at 640 pixels on MS COCO test-dev 2017 using an NVIDIA V100.At 1536 pixels, YOLOv5 achieved 55.8% AP.
  • YOLOv6 architecture: YOLOv6 introduced CSPDarknet, an FPN, and separated classification and box-regression heads for industrial real-time detection.The FPN expanded feature scales, while head separation added layers that segregated classification and regression functions.
  • YOLOv6 performance: 52.5% AP and 70% AP50 were reported for YOLOv6L at about 50 FPS on an NVIDIA Tesla T4 using MS COCO test-dev 2017.YOLOv6nano was described as the smallest and fastest variant, whereas YOLOv6small prioritized accuracy for smaller objects.

4.7 YOLOv7

YOLOv7 advanced efficiency and scalability through E-ELAN, scalable concatenation, model scaling, and training refinements. YOLOv8 then shifted to anchor-free prediction and reported higher AP and faster processing than YOLOv5 at the stated evaluation settings.

  • YOLOv7 architecture: YOLOv7 used E-ELAN and scalable concatenation to manage gradient paths, improve convergence, and produce models with varied computational scales.Its scaling strategy adjusted block depth and width uniformly while mitigating hardware-resource consumption.
  • YOLOv7 performance: YOLOv7’s reported speed enhancements ranged from 5 FPS to 160 FPS, reflecting its focus on efficiency and scalability.The model also incorporated bag-of-freebies techniques and re-parameterized convolution concepts.
  • YOLOv7 training: YOLOv7 used coarse label assignment for an auxiliary head and fine label assignment for the lead head, which produced the final output.The auxiliary head contributed to training, while the lead head generated predictions.
  • YOLOv7 evaluation: 55.9% AP and 73.5% AP50 were reported for YOLOv7E6 on MS COCO test-dev 2017.Batch-normalization statistics were amalgamated into convolutional weights and bias during inference.
  • YOLOv8 architecture: YOLOv8 adopted anchor-free center prediction instead of predetermined anchor boxes, addressing mismatches between anchors and custom dataset distributions.This represented a methodological departure from traditional anchor-based detection.
  • YOLOv8 evaluation: 53.9% AP and 280 FPS were reported for YOLOv8x at 640 pixels using an NVIDIA A100 with TensorRT, compared with YOLOv5’s 50.7% AP at the same image size.YOLOv8 was offered in five variants for different accuracy and computational requirements.

4.9 YOLOv9

YOLOv9 introduces PGI and GELAN to improve information flow, efficiency, and deployment flexibility, while reported results show a 0.6% AP improvement over YOLOv8 on MS COCO. The review then situates YOLOv10 as a further real-time detection advance centered on NMS-free inference and performance-efficiency trade-offs.

  • YOLOv9 innovations: YOLOv9 introduces Programmable Gradient Information (PGI) to address information bottlenecks and support deep supervision in lightweight architectures.
  • YOLOv9 innovations: GELAN is designed as an efficient, lightweight architecture suitable for varied computational configurations and resource-constrained edge devices.
  • YOLOv9 results: 0.6% AP improvement on MS COCO was reported for YOLOv9 compared with YOLOv8, alongside parameter reduction and computational-efficiency gains.
  • YOLOv10 innovations: YOLOv10 removes NMS during post-processing through dual label assignments and adds lightweight heads, decoupled downsampling, and rank-guided blocks.
  • YOLOv10 results: YOLOv10 reports lower latency and model size than YOLOv9 while maintaining competitive or superior detection accuracy, including on COCO.

5 Agricultural Applications of YOLO

Agricultural studies apply YOLO variants to weed detection, crop identification, disease detection, vehicle tracking, and animal monitoring. Reported results demonstrate strong accuracy and real-time potential, but field conditions, sensing hardware, dataset scope, and generalization remain recurring constraints.

  • Weed Detection Using YOLO: YOLO-based weed detection supports real-time identification for targeted agricultural management, including aerial spraying and weed control.
  • Weed Detection Using YOLO: 93.81% mean average precision and 0.94 F1 score were achieved by YOLO-WEED for UAV video weed detection in green onion fields.
  • Weed Detection Using YOLO: Weed-detection studies report limitations involving controlled environments, field validation, model performance, regional datasets, environmental conditions, and commercial sensing technology.
  • Crop Detection via YOLO: YOLO variants have been used for crop and fruit detection, including oil-palm objects, citrus fruit, wheat-ear Fusarium head blight, and agricultural vehicles.
  • Crop Detection via YOLO: 98.68% mean average precision and 0.97 F1 score were reported for YOLO-P in oil-palm plantations, with a lightweight 76 MB design.
  • Disease Detection in Agriculture Using YOLO: 93.69% accuracy was reported for a lightweight YOLOv4-MobileNet model detecting wheat-ear Fusarium head blight, with remaining issues for small objects, complex backgrounds, and cross-platform generalization.
  • Animal Tracking with YOLO: YOLO-based animal applications include pig counting, breeder-behavior analysis, and red-fox monitoring, but sample size, camera placement, rare poses, and training-data coverage constrain some studies.

6 Discussion

YOLO variants show broad utility in agricultural applications, combining real-time detection with accuracy across precision farming, weed management, and crop monitoring. Their adaptability supports crop-health assessment, yield estimation, and resource optimization, although small-object detection remains challenging.

  • Overall Discussion: The diverse YOLO models used across agricultural applications show the architecture’s adaptability and consistently high accuracy across multiple tasks.The reviewed applications include weed detection, crop monitoring, and other agricultural uses.
  • Real-Time Precision Agriculture: YOLOv3, YOLOv4, and YOLOv5 demonstrate swift and accurate object detection for real-time precision agriculture.Their real-time capability supports timely decisions and more efficient resource allocation.
  • Weed Detection and Management: YOLO-based systems enable real-time weed identification and localization for targeted and efficient weed control.YOLO-WEED and YOLOv4-weeds indicate potential for wider adoption despite resolution limitations with smaller objects.
  • Crop Detection and Monitoring: YOLO variants support accurate crop identification and delineation across large-scale fields for monitoring crop health, estimating yields, and optimizing resources.Specialized models including CitrusYOLO and YOLO-P illustrate adaptation to different crops and agricultural settings.

7 Challenges in YOLO-based Agricultural Applications

YOLO-based agricultural systems face limitations involving dataset specificity, deployment hardware, environmental variability, and detection of small objects. Addressing these constraints requires broader data, efficient deployment methods, environmental robustness, and improved feature extraction or multi-scale detection.

  • Research Priorities: Broad adoption requires more inclusive datasets, edge-device optimization, and models resilient to changing agricultural environments.These priorities directly address dataset specificity, hardware limitations, and environmental variability.
  • Dataset Specificity: Region- and crop-specific datasets can limit YOLO models’ generalizability across diverse agricultural landscapes.The review calls for datasets covering varied crops, growth stages, and environmental conditions, alongside transfer-learning efforts.
  • Hardware Limitations: Hardware constraints complicate real-world deployment of YOLO systems on edge devices and embedded precision-farming equipment.Model optimization, quantization, and lightweight architectures are proposed to improve accessibility in resource-limited settings.
  • Environmental Variability: Variations in illumination, weather, and terrain can challenge reliable YOLO detection in dynamic agricultural environments.The review identifies environmental adaptability during training as necessary for robust performance under diverse conditions.
  • Small-Object Detection: YOLO variants may struggle to detect diminutive weeds or diseases because small objects are difficult to distinguish within images.Potential responses include improved feature extraction, attention mechanisms, multi-scale detection, and ensemble techniques.

8 Future Directions and Opportunities

Future work emphasizes richer sensing, interpretable predictions, adaptive learning, and collaboration between farmers and AI systems. Together, these directions target more robust, trusted, and context-aware YOLO applications in precision agriculture.

  • Multi-Modal Integration: Combining RGB imagery with thermal or hyperspectral data could provide richer information for robust detection of crops, weeds, and diseases.The review identifies multi-modal fusion as a route toward more accurate agricultural detection.
  • Explainable AI: Explainability and interpretability are needed so farmers and stakeholders can understand the rationale behind YOLO predictions.The review prioritizes methods that explain how and why detections are made in complex agricultural environments.
  • Adaptive Systems: Real-time learning mechanisms could allow YOLO models to adapt and improve from continuous field feedback.This direction is linked to adaptive and self-improving precision-agriculture systems.
  • Human-AI Collaboration: Human-AI collaboration could combine farmer knowledge with AI-driven insights for more context-aware agricultural practices.The review presents collaboration as important for handling agricultural complexity and uncertainty.
  • Integrated Development: Progress depends on coordinated refinement of models alongside advances in data collection, hardware, and interpretability.The review connects these efforts with sustainable and efficient farming practices.

9 Conclusion

The survey concludes that YOLO variants offer versatile, real-time solutions for precision farming, crop monitoring, disease identification, and weed management. Their broader agricultural impact depends on addressing dataset, hardware, and environmental constraints while advancing multimodal, explainable, adaptive, and human-centered systems.

  • Core Conclusion: YOLO architectures demonstrate versatile and efficient applications across crop detection, disease identification, weed management, and precision farming.The conclusion characterizes these capabilities as relevant to long-standing agricultural challenges.
  • Remaining Challenges: Training-dataset specificity, hardware limitations, and environmental variability remain major hurdles for wider agricultural deployment.The review prioritizes inclusive datasets, edge-device optimization, and resilience to dynamic environments.
  • Future Opportunities: Multi-modal integration, explainability, real-time learning, and human-AI collaboration are identified as opportunities to enhance YOLO’s agricultural utility.These directions support more intelligent and context-aware precision farming.
  • Broader Implications: Continued collaboration among researchers, practitioners, and stakeholders is associated with advancing YOLO applications toward food security and environmentally conscious agriculture.The stated consequence remains within the review’s projected scope for future agricultural development.
Loading 2406.10139v1…