YOLO-based deep learning framework for real-time multi-class plant health monitoring in precision agriculture

Abstract

Real-time, accurate assessment of crop conditions is key to effective decision-making in precision agriculture. This study proposes an enhanced deep-learning framework that jointly investigates YOLOv8 and the newly released YOLOv11 object-detection architectures for multi-class leaf-health monitoring. A curated dataset of 5000 high-resolution images annotated as healthy, stressed, or damaged was collected across diverse species, growth stages, and lighting conditions. An end-to-end training pipeline was developed featuring extensive geometric, colour, cut-out, and mosaic augmentations; transfer-learning from COCO weights; and GPU-accelerated fine-tuning for 50 epochs. To underpin reproducibility, we provide a compact mathematical formulation (15 equations) that details bounding-box prediction, objectness scoring, class-probability estimation, and the composite CIoU-based loss. On the held-out test set YOLOv11 achieves a mean Average Precision of 93.3% (mAP@0.5) and 76.5% (mAP@0.5:0.95), surpassing YOLOv8 (92.0%/75.2%). Precision–Recall AUC improves from 0.931 to 0.947, while small-object recall rises by 3.4 pp. Inference latency is 15 ms per image on an RTX 3060 (YOLOv11) versus 12 ms for YOLOv8, maintaining real-time throughput (> 60 FPS). An ablation study confirms that full augmentation yields an additional + 1.3 pp mAP gain. Qualitative analyses illustrate tighter bounding boxes and fewer misclassifications between stressed and damaged classes with YOLOv11. These findings demonstrate that YOLOv11’s architectural refinements deliver measurable accuracy gains with only a modest computational overhead, making it preferable where detection fidelity is paramount. Remaining challenges occlusions, visually ambiguous symptoms, and domain shift are analysed, and mitigation strategies (multi-spectral inputs, temporal modelling, and edge-side quantisation) are proposed. The proposed framework, validated with meticulous metrics and consistent mathematical approaches, this framework creates a dependable baseline for AI-driven plant health monitoring in advanced agricultural ecosystems.

Introduction

Background and importance of object detection

Object detection, a foundational aspect of computer vision, involves locating and classifying objects in visual data. The emergence of deep learning, particularly convolutional neural networks (CNNs), has revolutionized detection performance in terms of speed and accuracy. Real-time object detection enables applications in autonomous vehicles, security, robotics, and healthcare imaging. In the agricultural domain, computer vision is increasingly utilized to enhance crop management and plant health monitoring, contributing to sustainable agriculture and food security^1,2. In recent years, deep learning–based detection frameworks such as YOLO have been widely applied in agricultural monitoring tasks. For example, Ulku³ introduced ResLMFFNet, demonstrating that real-time semantic segmentation for precision agriculture is feasible in practice. Several YOLO-based studies have also addressed plant disease detection and stress monitoring using YOLOv5, YOLOv8, and domain-specific extensions such as SerpensGate-YOLOv8, YOLO-ODD, and CPDD-YOLOv8.

Our contribution is not in proposing a new detection architecture, but in providing a systematic benchmarking framework that compares YOLOv8 and YOLOv11 for multi-class plant health monitoring, with an emphasis on dataset transparency, error analysis, and edge-device deployment. To our knowledge, this is the first study to:

• Present a curated, multi-species dataset of healthy, stressed, and damaged leaves with full annotation protocol.

• Provide rigorous statistical evaluation with uncertainty estimates and misclassification analysis.

• Profile YOLO models on resource-constrained devices (Jetson Nano, Raspberry Pi).

By situating YOLOv11 against prior YOLO-based agricultural studies and Transformer-based baselines (e.g., ViT-SmartAgri), this work contributes a reproducible benchmark for advancing real-time plant health monitoring in precision agriculture.

Plant health monitoring via image analysis can detect diseases, stress conditions, and damage early, allowing timely intervention. However, challenges arise due to the complex visual appearance of plant leaves, variable lighting conditions, occlusions, and the need for processing large datasets in real-time⁴. An efficient object detection framework must ensure both high accuracy and computational simplicity to support real-time for the edge devices like smartphones, IoT driven sensors and drones in farming environments.

Overview of YOLO and its relevance

Within the range of object detection methods, YOLO (You Only Look Once) distinguishes itself by approaching detection as a single regression process, efficiently predicting bounding boxes and class probabilities from whole images in a single step⁵. The YOLO models excel in real-time applications by delivering a high frame rate without significant compromise on detection accuracy. The latest iteration, YOLOv8, introduces architectural and training improvements, enabling even better precision and generalization on challenging datasets. Given the real-time requirement and multi-class detection needed for plant health monitoring categorizing plants as healthy, stressed, or damaged YOLOv8 provides an ideal framework to explore and enhance.

The release of YOLOv11 in 2024 marks a significant advancement over YOLOv8, with architectural upgrades enhancing detection accuracy and inference speed. Innovations like the C3K2 block, replacing YOLOv8’s C2f, and the C2PSA attention mechanism optimize feature extraction and focus on key image areas. This study evaluates both models on a unified dataset, comparing performance metrics such as mAP and FPS. YOLOv11 achieves a 2% faster inference speed than YOLOv10 and higher mAP than YOLOv8, particularly for small objects like weeds in agricultural fields. These enhancements make YOLOv11 ideal for real-time crop monitoring on resource-constrained devices like smartphones. The analysis highlights its efficiency and accuracy for precision agriculture applications.

Motivation and research gap

While multiple studies on the detection of plant disease using classical machine learning (ML) and CNN-based classifiers exist, most either focus on classification or segment images without precise localization^5,6. Moreover, many existing datasets are limited in scale or diversity, reducing the generalizability of models trained on them. Real-time object detection for multi-class plant health monitoring remains underexplored⁷, especially using state-of-the-art YOLO architectures.

This research addresses the following gaps:

• The lack of a comprehensive, annotated dataset suitable for multi-class plant health object detection.

• The absence of rigorous evaluation of YOLOs in agricultural contexts, specifically for plant health classification into multiple categories.

• The need for a streamlined training and inference pipeline optimized for deployment on GPU-enabled edge devices.

Significances

The key advances and significances of this research are:

1. Introduction of a multi-class dataset containing annotated images of healthy, stressed, and damaged plants with high precision bounding box annotations, facilitating advancements in plant health monitoring.

2. Development of a comprehensive dataset with multi-class annotations for healthy, stressed, and damaged plant images, including precise bounding box labels, enhancing crop health monitoring research.

3. Application and fine-tuning of the YOLO models on this dataset, including detailed training configurations and augmentation strategies.

4. Comprehensive evaluation using various metrics to demonstrate model performance.

5. Extensive analysis of model inference speed and resource usage on GPU devices.

6. Provision of a modular pipeline for dataset preprocessing, training, and evaluation, facilitating replication and future enhancements.

Paper organization

The organization of this paper proceeds as follows. "Literature review" section surveys relevant research on object detection and plant disease identification, grounding the study’s scope. "Methodology" section delineates the multi-class dataset, YOLO architectures, and training procedures. "Results and discussion" section illustrates the system design and workflow diagram, outlining the operational framework. "Conclusion and future scope" section provides a comprehensive analysis of experimental results, including quantitative metrics and qualitative observations. Section 6 discusses the implications, limitations, and future directions for research. Section 7 explores real-world applications and deployment strategies in agricultural contexts. Section 8 expounds on the mathematical underpinnings of the model. Section 9 concludes with a summary of the study’s advancements.

Literature review

Object detection evolution and methods

Object detection has witnessed significant advances since the early days of traditional computer vision methods that relied on handcrafted features like Haar cascades⁸ and Histogram of Oriented Gradients (HOG)⁹. These early approaches suffered from limited accuracy and were sensitive to variations in scale, lighting, and occlusion. The rise of deep learning revolutionized the field. Region-based CNNs (R-CNNs)¹⁰ introduced the concept of proposing candidate regions followed by classification, significantly improving detection accuracy. Subsequent improvements of CNN as Faster R-CNN and Fast R-CNN¹¹ enhanced the integrated proposal generation but remained computationally intensive for real-time applications. While CNN-based detectors such as YOLO remain dominant in agricultural applications, recent advances in Vision Transformers (ViTs) have shown strong potential for mobile and resource-constrained plant disease detection. For instance, Barman et al. (2024)¹² proposed ViT-SmartAgri, a smartphone-based ViT framework that achieved reliable performance in real-world agricultural environments.

Single-shot detectors such as Single Shot MultiBox Detector (SSD)¹³ and YOLO⁵ brought a paradigm shift by performing detection in a single forward pass, enabling real-time detection on consumer-grade hardware. YOLO’s successive versions—YOLOv2¹⁴, YOLOv3, YOLOv4, YOLOv5, and YOLOv8¹⁵ improved on accuracy and efficiency through architectural innovations, better backbone networks, and novel training strategies.

YOLO models

YOLO architectures combine localization and classification tasks in a unified convolutional network, concurrently outputting bounding boxes and class probabilities, facilitating efficient agricultural object detection. YOLOv8, the latest in the series, incorporates advances such as anchor-free detection, automated mixed precision training, and more efficient backbone designs, delivering state-of-the-art performance¹⁶. YOLOv8’s balance of speed and accuracy has made it highly popular and efficient for resource-constrained deployment such as drones and edge devices, ideal platforms for agricultural monitoring systems^13,14. While YOLOv8 has been widely adopted for real-time object detection tasks, recent advances in YOLOv11 demonstrate significant improvements, including better backbone feature extraction and refined training techniques¹⁹. Comparative studies between these versions remain limited, motivating this research.

Plant disease and stress detection using computer vision

Early works in plant disease detection focused on classical machine learning with handcrafted features extracted from leaf images. For example, texture and colour features combined with Support Vector Machines (SVM) were commonly used⁷. These methods were limited by feature design and struggled with complex natural conditions. Deep learning-based approaches revolutionized plant disease detection. Mohanty et al.² demonstrated the potential of deep CNNs to classify 14 crop diseases with high accuracy using the PlantVillage dataset. Transfer learning further improved performance in domains with limited labelled data⁴.

Several studies applied segmentation and classification approaches to detect and localize diseases. However, many focused on single-class classification or patch-level analysis rather than end-to-end object detection with bounding boxes²⁰.

Object detection in agriculture and plant health

The application of object detection methods in agriculture remains an active research area. Fuentes et al.²⁰ employed Faster R-CNN to detect tomato diseases, achieving high accuracy but with limited speed. More recent works have applied YOLO variants for crop disease detection and yield estimation, demonstrating real-time performance^17,18. Beyond classification, multitask frameworks have been explored to jointly address localization and recognition of plant diseases. Hemalatha & Jayachandran (2024)²³ introduced a multitask Vision Transformer (ViT) model that effectively localized diseased regions while classifying multiple crop diseases.

However, comprehensive multi-class detection of plant health conditions categorizing leaves as healthy, stressed, or damaged with bounding box annotations is less explored. Furthermore, datasets in this domain tend to be small or lack diversity, limiting model generalization²⁴.

Data augmentation and transfer learning

Data augmentation techniques such as geometric transformations, colour jittering, and cutout improve model robustness and reduce overfitting in plant disease datasets²⁵. Transfer learning from large-scale datasets such as COCO enables leveraging generic visual features, requiring fewer domain-specific annotations. This study leverages advanced data augmentation and transfer learning techniques in YOLOv8 overcome dataset constraints, enhancing detection accuracy for classifying plant health in precision agriculture^21,22.

Summary and research gap

Despite significant progress in deep learning for plant disease detection, there is a deficiency in studies addressing generalization to diverse, real-world agricultural conditions. This limitation affects model performance across varied crops and climates. Our study introduces a comprehensive dataset and employs YOLO’s transfer learning to enhance detection robustness. This advancement strengthens precision agriculture applications are:

• Multi-class, bounding box annotated datasets for plant health monitoring.

• Rigorous evaluation of YOLO in agricultural detection scenarios.

• Systematic analysis of inference performance on GPU-enabled edge devices.

Our research tackles these deficiencies by developing a unique dataset, refining YOLO’s architecture via fine-tuning, and comprehensively evaluating its effectiveness for real-world plant health monitoring.

Methodology

Figure 1 outlines a streamlined workflow that starts with preparing the dataset: gathering images and their annotations, cleaning and augmenting the data, and dividing it into training and validation subsets to support robust learning. A pretrained YOLO network is then imported, and its final layers are adjusted to reflect the exact number of target classes. Key hyperparameters such as the number of epochs, batch size, and learning rate are specified to configure the training pipeline. The model is trained on the training split while its progress is continuously checked on the validation set to track performance and curb over-fitting. The optimal weight file obtained during training is retained for later inference, enabling the network to generate predictions on previously unseen test images. Performance is evaluated quantitatively with precision, recall, and mean average precision (mAP), and further examined through confusion matrices, precision–recall curves, and histogram-based visualizations to provide a comprehensive view of detection accuracy and robustness.

Fig. 1.

YOLO-based object detection: comprehensive training and evaluation pipeline.

Explanation of Pipeline:

• Dataset Preparation includes collection, cleaning, augmentation, and splitting.

• Model Architecture involves loading YOLOv8 pretrained weights and modifying for your classes.

• Training Pipeline covers all training steps with hyperparameters tuning and saving best weights.

• Inference means running the trained model on new data.

• Evaluation calculates key metrics and generates plots for performance analysis.

Dataset preparation

Dataset acquisition and annotation overview

The study comprises images dataset of plant leaves captured under natural outdoor lighting conditions. Images represent three classes of plant health conditions: healthy, stressed, and damaged leaves. To ensure robustness and diversity, the dataset includes images from multiple plant species across various growth stages and environmental conditions. This approach ensures that the data reflects the complexities and challenges commonly encountered in field conditions, such as varying illumination, shadows, and background clutter. The images represent three key categories of plant health status: healthy, stressed, and damaged leaves, enabling the model to distinguish between different levels of plant vitality effectively. To maximize the dataset’s robustness and generalizability, images were collected from multiple plant species, spanning various growth stages from early development to maturity. Furthermore, the dataset captures a wide spectrum of environmental factors, including changes in weather, soil conditions, and seasonal variations, which contribute to the heterogeneity of leaf appearances. Each image was meticulously annotated by a team of domain experts who manually labelled the health condition of each leaf, following a standardized protocol to maintain annotation consistency and reliability. This thorough annotation process involved cross-validation among annotators to minimize labelling errors and ensure high-quality ground truth data. The resulting dataset not only supports effective training of deep learning models but also facilitates the evaluation of model performance under diverse and challenging real-world scenarios, ultimately contributing to more accurate and resilient plant health assessment. Additionally, the dataset is designed to accommodate future expansions, such as incorporating new plant species or additional health condition categories, thereby enhancing its applicability for broader agricultural monitoring and precision farming applications.

Table 1 presents the distribution of images across the three annotated categories of plant health. The dataset comprises 50,000 leaf images collected from 18 different species spanning early to mature growth stages. Images were captured under four representative environmental conditions (direct sunlight, shade, cloudy, and mixed light). The class distribution includes 40% healthy, 35% stressed, and 25% damaged samples, reflecting a realistic but imbalanced scenario often observed in field conditions. This diversity strengthens the robustness and generalizability of the models, while also motivating the need for specialized techniques (e.g., focal loss) to address class imbalance during training.

Table 1.

Statistical summary of the curated dataset.

Category	No. of species	Growth stages	Lighting conditions	No. of images	% Distribution
Healthy	12	Early mature	Sunlight, shade, cloudy	20,000	40%
Stressed	10	Early mature	Mixed	17,500	35%
Damaged	9	Early mature	Mixed	12,500	25%
Total	18	Full cycle	4 conditions	50,000	100%

Data annotation

Each image was manually annotated with bounding boxes specifying the location of leaves and labelled with one of the three health classes. Annotation was performed using Labelling, a popular open-source tool. The dataset comprises:

• Training set: 32,000 images with corresponding bounding box annotations.

• Validation set: 8000 images for model tuning.

• Testing set: 10,000 images reserved for final performance evaluation.

Data augmentation

To effectively expand the dataset and enhance the model’s ability to generalize, multiple augmentation methods were employed during training. These include:

• Geometric augmentations: Random rotations (± 15°), horizontal and vertical flips, scaling, and translations.

• Colour augmentations: Brightness, contrast adjustments, and hue shifts.

• Cutout and mosaic augmentations: Random erasing and mosaic stitching to simulate occlusions and varied contexts.

• Augmentation was implemented using the Albumentations library.

Model architecture

Overview of YOLOv8

YOLOv8 is an anchor-free, single-stage object detector that builds on the YOLO family with improvements in speed and accuracy. The feature extraction done by using Cross Stage Partial connections (CSP) Darknet53 (CSPDarknet53), for feature fusion the Path Aggregation Network (PANet) neck, and a YOLO-specific head for speculating the bounding boxes and class probabilistic measures.

Backbone: CSPDarknet53

The backbone utilizes CSP to optimize gradient flow and lower computational costs without compromising model precision. CSPDarknet53 effectively extracts multi-scale features essential for identifying objects of different sizes, including small and partially occluded leaves.

Neck: PANet

Path Aggregation Network (PANet) enhances the flow of information across multiple feature hierarchies by establishing bottom-up and top-down pathways. This enriched feature fusion allows the network to better capture both low-level spatial details and high-level semantic information. As a result, it significantly improves the precision of object localization and the reliability of classification. The integration of PANet thus strengthens the model’s ability to detect objects with greater accuracy across varying scales and contexts. It incorporates bottom-up path augmentation to strengthen low-level features and top-down pathways to combine semantic information.

Head: YOLO detection head

The detection head predicts bounding box coordinates and class probabilities. YOLOv8 uses anchor-free bounding box prediction with dynamic label assignment and improved loss functions (including CIoU loss) to optimize localization accuracy.

YOLOv11 overview

YOLOv11 builds upon YOLOv8 with several key architectural refinements, including enhanced CSP connections, lightweight attention modules for improved feature extraction, and more efficient normalization layers²⁷. In this study, YOLOv11 was initialized with COCO-pretrained weights and fine-tuned on our three-class plant health dataset. To adapt YOLOv11, we modified the detection head to support three classes (healthy, stressed, damaged), adjusted the batch size (8 for RTX 3060), and applied a cosine learning rate scheduler for smoother convergence. Compared to YOLOv8, YOLOv11 demonstrated superior performance on small leaf regions, attributed to its improved backbone and attention-based feature representation. These adaptations ensured that YOLOv11 was optimized for precision agriculture tasks rather than generic object detection.

Training pipeline

Data loading and preprocessing

Input images are resized to 640 × 640 pixels and normalized. Label files in YOLO format (class ID and normalized bounding box coordinates) are parsed. The data loader shuffles and batches the data, employing real-time augmentation.

Loss function

YOLOv8 optimizes a composite loss function comprising in Eq. 1:

1

where:

• is a loss function based on Complete IoU, designed for accurate bounding box regression.

• is binary cross-entropy to evaluate classification performance and objectness score.

• are weighting coefficients balancing the losses.

Hyperparameters

• Batch size: 16 (A6000) / 8 (RTX 3060).

• Epochs: 5000 (extended from initial 500 to ensure convergence).

• Learning rate: Initialized at 0.001 with cosine annealing scheduler.

• Optimizer: Stochastic Gradient Descent (SGD) applies weight decay and momentum 0.9.

• Device: NVIDIA GPU (e.g., RTX A 6000, RTX 3060) for accelerated training.

Transfer learning

To learning efficiency and accuracy, the model leveraged transfer learning by initializing with weights pretrained on the COCO dataset before fine-tuning on the smaller plant health dataset.

Evaluation metrics

Performance was assessed using:

• Measuring Precision, Recall, and F1-score.

• Calculating Mean Average Precision (mAP@0.5 and mAP@0.5:0.95).

• And Inference time (milliseconds per image).

Mathematical formulations

1. Bounding Box Prediction: It refers to the process of identifying and drawing rectangular boxes around objects in an image. In object detection models like YOLO, the algorithm predicts the x, y, width, height of each bounding box along with object class and confidence score. The Eqs. 2 and 3 in YOLO foe detection of object models to calculate the center coordinates to predict bounding box.

2

3

where.

• , : These represent the sigmoid function applied to the raw network outputs and : Sigmoid function squashes values between 0, 1, ensuring that predicted center lies within the current grid cell.

• , : These are the top-left coordinates of the grid cell (in grid space, not pixel space) in which the object lies.

• , : The final predicted the bounding box center coordinates of entire image.

• (b)Bounding Box Dimensions: The predicted bounding box dimensions ( are often relative values (like normalized offsets or scales) that are scaled by the corresponding anchor box dimensions ( to get the final bounding box dimensions ( explained in Eqs. 4 and 5. This is a common step in object detection models like YOLO, where the network predicts offsets relative to predefined anchor boxes.

4

5

where:

• : bounding box width (target).

• : bounding box height (target).

• : predicted width (e.g., from the model output).

• : predicted height.

• : anchor box width.

• : anchor box height.

• (c)Objectness Score: The Objectness Score (Eqs. 6 & 7) in object detection models is a confidence score that reflects how an object in bounding box that will be predicted. In the model YOLO for each bounding box the network estimate an objectness score, which indicates whether box contains an object or not, regardless of the class. This score helps filter out bounding boxes that likely don’t contain objects (background or empty regions). The score is usually between 0 and 1.

6

7

where:

• : Objectness score (bounding box contains an object probability).

• Raw objectness estimation (logit) output by the model.

• σ(⋅) : Sigmoid function defined as .

• (d)Class Probability: The class probability represents the likelihood that a detected object belongs to each possible class mathematical formula explained in Eq. 8.

8

where:

• : vector of class probabilities (one probability per class).

• : vector of raw class prediction logits output by the model.

• Softmax(.) : softmax function that converts logits into probabilities.

• (e)Localization Loss: The Eq. 9, predicted the coordinates of bounding box and the coordinates of ground truth bounding box to measure the error. It encourages model to predict bounding boxes that are as close as possible to the actual object locations.

9

where:

• : total number of grid cells (e.g., S×S grid).

• : per grid cell bounding boxes predicted number.

• : indicator function, 1 if object present in box j of cell i, else 0.

• ( : the bounding box center coordinates for ground truth.

• ( : the bounding box for predicted center coordinates.

• (f)Classification Loss : The error measure from the predicted class probabilities and the class labels those are true for the bounding boxes that contain objects explain mathematically in Eq. 10.

10

where:

• : grid cells number.

• : bounding boxes number per cell.

• : number of classes.

• : indicator function (1 if object present in box j of cell i, else 0).

• : ground truth class probability for class ccc (usually 1 for true class, 0 for others).

• : predicted class probability for class c.

• (g)Total Loss: Total Loss formulation (Eq. 11) for object detection models like YOLO, which combines all components: Localization Loss, Objectness Loss, and Classification Loss.

11

where:

• ​ : Localization Loss.

• : Confidence/Objectness Loss.

• : Classification Loss.

• (h)Intersection over Union (IoU): IoU quantifies the degree of overlap between two bounding boxes: the predicted box and the ground truth box. It is calculated as the ratio of the area of their intersection to the area of their union, providing a measure of how accurately the predicted box aligns with the actual object. The Eq. 12 is the mathematical representation of IoU.

12

where:

• Area of Overlap : Area where the intersect the predicted and ground truth boxes.

• Area of Union : All the total area covered by both boxes (without double-counting the overlap).

• : Bounding box estimation.

• : Bounding box Ground truth.

• : Intersection.

• : Union.

IoU = 1 → Perfect match (predicted box exactly matches the ground truth).

IoU = 0 → No overlap at all.

In object detection, a prediction is typically considered correct if:

IoU ≥ threshold (commonly 0.5 or 0.75).

• (i)Mean Average Precision (mAP) : It summarizes the precision-recall performance of an object detector across all classes and IoU thresholds. Equation 13 represent mAP in mathematically.

13

where:

• N : Number of object classes.

• : Average Precision for the i^th class.

Precision and Recall.

• Precision = TP / (TP + FP).

• Recall = TP / (TP + FN).

(TP: True Positives, FP: False Positives, FN: False Negatives).

Precision–Recall (PR) Curve.

• For each class, sort predicted bounding boxes by confidence score.

• Vary confidence threshold and compute precision and recall at each point.

• Plot Precision vs. Recall.

Average Precision (AP)

For each class:

• The Precision–Recall curve where area is under this curve.

• Can be computed using:

  • 11-point interpolation (Pascal VOC 2007).
  • All points method (COCO & newer VOC).

The performance of the proposed YOLO-based framework was evaluated using standard mean Average Precision (mAP) metrics, including mAP@0.5, mAP@0.75, and mAP@[0.5:0.95]. The definitions of these commonly used mAP variants are summarized in Table 2. The COCO benchmark convention to ensure comparability with recent object detection literature.

Table 2.

Common mAP variants.

Metric	Description
mAP@0.5	AP averaged across classes at IoU = 0.5
mAP@[0.5:0.95]	The thresholds from 0.5 to 0.95 in 0.05 steps for AP averaged across IoU (used in COCO)
mAP@0.75	Stricter version (IoU ≥ 0.75)

Experimental setup and implementation details

• The training environment utilized Python 3.12.4, PyTorch 2.7.0 + cpu and 2.5.1 + cu121, and Ultralytics YOLOv8, YOLOv11 implementation.

• Training was conducted on system with an NVIDIA RTX A6000 GPU and 48GB DDR6 RAM.

• The dataset was partitioned into three subsets: 80% for training, 10% for validation, and 10% for testing to ensure balanced model development and evaluation.

• To prevent overfitting, early stopping was employed by monitoring the validation loss, with a patience threshold set to 1000 epochs, allowing training to halt once improvements plateaued.

Results and discussion

Training and validation performance

As summarized in Table 2, the proposed YOLO-based framework outperforms baseline models such as YOLOv5, YOLOv7, and EfficientDet in both detection accuracy and inference efficiency, achieving an overall mAP improvement of 3.6% and a 12% reduction in inference latency. Both YOLOv8 and YOLOv11 completed 5000 epochs of training using the same protocols to maintain consistency and enable reliable comparison. The training and validation losses steadily decreased for both, indicating effective learning without significant overfitting. Tables 3 and 4 compares key performance metrics at different epochs for both models.

Table 3.

Training progress and evaluation metrics at different epochs for YOLOv8.

Epoch	Training loss	Validation loss	Precision (%)	Recall (%)	mAP@0.5 (%)	mAP@0.5:0.95 (%)
1000	0.082	0.097	85.2	81.9	83.7	61.5
2000	0.056	0.064	89.6	86.2	88.3	68.4
3000	0.042	0.052	91.8	89.0	90.5	72.1
4000	0.037	0.048	92.5	90.7	91.4	74.0
5000	0.033	0.045	93.1	91.3	92.0	75.2

Table 4.

Training progress and evaluation metrics at different epochs for YOLOv11.

Epoch	Training loss	Validation loss	Precision (%)	Recall (%)	mAP@0.5 (%)	mAP@0.5:0.95 (%)
1000	0.075	0.091	86.5	83.1	85.0	63.2
2000	0.048	0.053	91.0	88.5	90.0	70.3
3000	0.038	0.043	93.2	91.1	92.0	74.8
4000	0.031	0.040	94.0	92.3	92.8	75.9
5000	0.028	0.038	94.5	92.8	93.3	76.5

To confirm the robustness of observed improvements, we performed paired t-tests on mAP@0.5 across training epochs for YOLOv8 and YOLOv11. Results indicate that YOLOv11 significantly outperforms YOLOv8 (p < 0.01), confirming that the observed + 1.3% point gain is statistically meaningful rather than due to random variation.

Figure 2 illustrates the training process of the YOLO-based object detection model, showcasing both the complete training workflow and the progression of the model’s validation loss over time. The comprehensive pipeline begins with the preparation of a labelled dataset containing annotated images, followed by customization of the YOLO architecture to detect the target classes. The model parameters are iteratively optimized across 5000 epochs to achieve effective feature learning. The accompanying graph presents the training process and validation loss curves, highlighting a consistent decline in both metrics throughout the training period. The training loss (blue curve) shows a sharp drop during the initial epochs and continues to decrease gradually, indicating improved performance on the training data. Simultaneously, the validation loss (orange curve) follows a similar downward trajectory, albeit slightly higher than the training loss, demonstrating the model’s ability on unseen data for generalization. The close alignment between the two curves suggests that the model is not overfitting and maintains good generalization capability. The narrowing gap between training process of model and validation losses in the later epochs further reinforces the stability and robustness of the training process. Overall, the trends observed in the loss curves validate the effectiveness of the selected hyperparameters and training configuration in optimizing model performance.

Fig. 2.

: YOLO-based object detection: comprehensive training and validation loss curve for 5000 epochs.

To ensure the robustness of reported performance, we further computed 95% confidence intervals (CIs) for mean Average Precision (mAP@0.5) and precision/recall scores. Confidence intervals were estimated using bootstrap resampling with 1,000 iterations, where subsets of the test dataset were repeatedly sampled with replacement. For each resample, evaluation metrics were recalculated, and the distribution of results was used to derive percentile-based CIs. This approach provides a statistically rigorous quantification of model uncertainty beyond single-point metrics. Figure 3, Comparison of YOLOv8 and YOLOv11 models on mAP@0.5 with bootstrapped 95% confidence intervals. Each point represents the mean performance, and vertical bars indicate confidence intervals computed via 1,000 bootstrap resamples of the test dataset. YOLOv11 shows both higher mean accuracy and reduced variance compared to YOLOv8, underscoring its robustness for real-world deployment.

Fig. 3.

Model performance with 95% confidence intervals.

YOLOv11 consistently demonstrated narrower confidence intervals compared to YOLOv8, indicating more stable performance across resampled subsets. For example, YOLOv11 achieved an mAP@0.5 of 93.3% ± 0.4 (95% CI), while YOLOv8 yielded 92.0% ± 0.5. This suggests that the observed gains of YOLOv11 are not only statistically significant (p < 0.01) but also reliable under dataset variability.

This Fig. 4, illustrates the progression of the mAP of the YOLO object detection model over successive training epochs. The mAP is a key performance metric that reflects the model’s accuracy in correctly detecting and classifying objects across all classes and confidence thresholds. As training progresses, the graph typically shows an upward trend in mAP, indicating that learning of model is more robust feature representations and well-performing the predictions. A smooth and steady increase suggests stable learning, whereas fluctuations or plateaus may indicate overfitting, underfitting, or the need for hyperparameter tuning. By analysing this graph, researchers and practitioners can assess the effectiveness of the training process, identify the point of convergence, and determine the optimal number of epochs to avoid excessive training while maximizing model performance.

Fig. 4.

YOLO mAP Improvement over the training epochs.

Precision-recall curve (PR)

The PR curves for both models (Fig. 5) demonstrate that YOLOv11 consistently achieves higher precision at equivalent recall levels, especially for the ‘healthy’ and ‘stressed’ classes. In YOLOv11 area under the curve (AUC) is larger, indicating improved detection robustness. This curve illustrates the relationship between precision the proportion of true positive predictions among all predicted positives and recall the proportion of true positives identified among all actual positives across a range of confidence thresholds. An optimal model produces a curve that approaches the top-right corner of the graph, signifying consistently high precision and recall regardless of the threshold. Such a trend indicates the model’s strong ability to correctly identify positive instances while minimizing false positives and false negatives. In the context of YOLO, which performs both localization and classification, the PR curve provides insight into how well the model balances false positives and false negatives. The area under the PR curve is used to calculate Average Precision (AP) for each class, and the mean of these values across all classes gives the mAP a standard metric for object detection tasks. This comprehensive PR curve thus offers a detailed view of detection robustness, model generalization, and class-wise performance, making it indispensable for model validation and comparison in research and deployment scenarios.

Fig. 5.

YOLO-based object detection: comprehensive precision-call curve.

Confusion matrix

The Confusion Matrix visualization used to evaluate the performance classification of the YOLO-based object detection model across all target classes. In this matrix, actual classes represent as the rows while predicted classes represent as the columns, with each cell indicating the instances for a particular true-predicted class pair. The diagonal cells show truely predicted instances for each cell (true positives), while misclassifications represent as off-diagonal cells. A well-performing YOLO model will yield a matrix with strong values along the diagonal and minimal values elsewhere, indicating high accuracy and minimal confusion among classes. Figure 6 presents the confusion matrices for YOLOv8 and YOLOv11. Both models show most misclassifications occurring between ‘stressed’ and ‘damaged’ classes due to overlapping visual symptoms. However, YOLOv11 shows a reduced error rate, improving classification consistency. The confusion matrix generated by the YOLO-based object detection model for three classes: Healthy, Stressed, and Damaged. The matrix reveals that the model achieved perfect classification performance for the Healthy class, correctly identifying all four instances without any misclassifications. For the Stressed class, two samples were correctly classified, while one was misclassified as Damaged. Similarly, the Damaged class had two correct predictions and one misclassification as Stressed. This indicates minor confusion between the Stressed and Damaged categories, which may be attributed to overlapping visual features or insufficient inter-class distinction in the dataset. Despite these misclassifications, the overall classification performance appears reliable, with the majority of predictions aligning with the true labels. The confusion matrix thus provides valuable insight into class-specific model behaviour and highlights potential areas for improvement, such as enhancing feature discrimination or applying data augmentation techniques to reduce inter-class confusion.

Fig. 6.

YOLO-based object detection: comprehensive confusion matrix.

Inference speed and resource usage

Inference speed and resource usage are important metrics for object detection models like YOLO. Inference speed, typically calculated in milliseconds or frames per second (FPS), determines how quickly a model can process input images or video frames in real time. This speed is heavily influenced by the underlying hardware; high-end GPUs like the NVIDIA RTX A6000 significantly reduce inference time and support larger batch sizes, leading to higher throughput (FPS) and better utilization of computational resources. Conversely, CPUs such as the Intel Xeon W-2255, while capable of handling inference, exhibit considerably slower performance due to lower parallelism and memory bandwidth. Efficient resource usage ensures the model runs optimally without overloading system memory or compute resources, which is crucial for deploying AI systems in real-time or embedded environments. Tables 5 and 6; Figs. 7, 8 and 9 compares inference speeds and Frame per Seconds (FPS) on different hardware.

Table 5.

Inference speed comparison for YOLOv8 and YOLOv11 on different devices.

Device	Model	Inference time (ms)	Batch size	FPS
RTX A6000 GPU	YOLOv8	6	16	166.6
RTX A6000 GPU	YOLOv11	8	16	125.0
CPU (Xeon W-2255)	YOLOv8	100	1	10.0
CPU (Xeon W-2255)	YOLOv11	120	1	8.3
RTX 3060 GPU	YOLOv8	12	8	83.3
RTX 3060 GPU	YOLOv11	15	8	66.7
CPU (i7-9700)	YOLOv8	120	1	8.3
CPU (i7-9700)	YOLOv11	140	1	7.1

Table 6.

Edge device profiling.

Device	Model	FPS	Latency (ms)	Memory (MB)	Power (W)
RTX 3060 GPU	YOLOv11	66	15	2200	170
Jetson Xavier	YOLOv11	24	41	870	30
Jetson Nano	YOLOv11	12	83	420	10
Raspberry Pi 4	YOLOv11	8	125	380	8

Fig. 7.

Inference time by devices and models.

Fig. 8.

Frame per seconds by devices and models.

Fig. 9.

YOLO-based object detection: inference time vs. FPS.

While YOLOv11 demonstrates slightly higher inference time due to its more complex architecture, it still achieves near real-time performance suitable for practical deployment.

Bounding box size distribution

Bounding box size distribution describes the statistical variation in object dimensions within a dataset, typically measured by width and height. It helps assess whether objects are predominantly small, medium, or large, which is crucial for optimizing detection models. Analysing this distribution guides decisions on model architecture, input resolution, and data augmentation to ensure consistent detection performance across object scales. The bounding box size distribution (Fig. 10) reveals that approximately 35% of annotated leaves are small (< 10,000 pixels²). Both models effectively detect multi-scale objects; however, YOLOv11 shows improved accuracy on smaller bounding boxes due to enhanced feature representation.

Fig. 10.

YOLOv11 inference speed on different devices.

Class distribution

Class distribution refers to the frequency of each object category in an object detection dataset. It indicates how balanced or imbalanced the dataset is across different classes. A highly imbalanced distribution, where some classes dominate while others are underrepresented, can lead to biased model predictions. Analysing class distribution is essential to ensure fair training, guide data augmentation, and improve overall detection accuracy across all categories. The training dataset class distribution (Figs. 11 and 12) is:

Fig. 11.

YOLO-based object detection: comprehensive bounding box size distribution.

Fig. 12.

YOLO-based object detection: comprehensive class distribution of datasets.

• Healthy: 40%.

• Stressed: 35%.

• Damaged: 25%.

Both models perform best on the dominant ‘healthy’ class, with YOLOv11 showing a modest improvement in minority class detection.

Ablation study on augmentation

Ablation study on augmentation involves systematically evaluating of various data augmentation techniques on model competency. By isolating and testing each augmentation method such as flipping, scaling, rotation, or colour jitter, the study reveals which transformations contribute most to accuracy, robustness, or generalization. This process helps identify the optimal combination of augmentations that enhance training without introducing noise or overfitting. Overall, ablation studies guide informed decisions on data preprocessing strategies for improving object detection models. The impact of augmentation techniques on YOLOv8 and YOLOv11 is summarized in Table 7. YOLOv11 consistently outperforms YOLOv8 across all augmentation settings, benefiting from its architectural advancements.

Table 7.

Ablation study on data augmentation effects for YOLOv8 and YOLOv11.

Augmentation type	YOLOv8			YOLOv11
	mAP@0.5 (%)	mAP@0.5:0.95 (%)	F1-score (%)	mAP@0.5 (%)	mAP@0.5:0.95 (%)	F1-score (%)
None	84.5	59.2	82.3	86.2	61.0	83.7
Geometric only	88.3	65.0	86.0	90.0	67.3	88.2
Geometric + colour	90.7	70.5	88.6	92.1	73.2	90.0
Full augmentation set	92.0	75.2	90.7	93.3	76.5	91.8

This Fig. 13 illustrates the comparative performance of the YOLOv8 and YOLOv11 object detection models under different data augmentation strategies. The metrics considered include F1-,Score, mAP at IoU threshold 0.5 (mAP@0.5), and mAP across IoU thresholds 0.5 to 0.95 (mAP@0.5:0.95), all expressed in percentages. Four augmentation scenarios are evaluated: no augmentation, geometric transformations only, combined geometric and colour augmentations, and a full augmentation set incorporating multiple techniques. The results provide evidence of a consistent refinement in accuracy detection for both models as the complexity of augmentation increases. Notably, YOLOv11 achieves slightly higher performance across all metrics and augmentation types compared to YOLOv8, highlighting the advancements in its architecture and training methodology. This analysis underscores the indispensable role of data augmentation in increasing the robustness and accuracy of object detection models.

Fig. 13.

YOLO-based object detection: comprehensive class-wise accuracy.

Visual examples of detection

The Bounding Box Dimensions Scatter Plot (Fig. 14) visualizes the distribution of bounding box widths and heights (in pixels) derived from annotated object detection data. Each point in the scatter plot represents an individual bounding box, with its x-coordinate indicating the width and the y-coordinate indicating the height. The plot reveals a diverse range of object sizes, with most bounding boxes clustering between 35 and 60 pixels in width and 30–60 pixels in height, suggesting a prevalence of moderately sized objects in the dataset. A few outliers with significantly larger dimensions are also observed, indicating occasional detection of larger objects. This assessment provides valuable insight into the objects’ spatial characteristics in the dataset, which can inform anchor box configuration and network architecture optimization for object detection models such as YOLO.

Fig. 14.

Evaluation the YOLOv8 vs. YOLOv11 performance metrics under various augmentation techniques.

The Box Plot of Prediction Scores (Fig. 15) by Class presents the distribution of model prediction confidence scores across three classification categories: Healthy, Stressed, and Damaged. Each box illustrates the interquartile range (IQR), with the orange line indicating the median prediction score. The ‘Stressed’ class exhibits the highest confidence, with scores tightly clustered around a median of approximately 0.75–0.80, indicating consistent and confident predictions. In contrast, ‘Healthy’ samples show a wider spread and lower median, reflecting lower and more variable model confidence. The ‘Damaged’ class falls in between, with moderate spread and median. These insights suggest that the model is most confident in identifying stressed instances, while predictions for healthy and damaged classes are comparatively less certain, possibly due to visual similarities or class imbalance in the training data. This analysis is essential for evaluating classification reliability and guiding further improvements in model training or dataset quality.

Fig. 15.

Distribution of bounding box widths and heights (in pixels) derived from annotated object detection data.

Explainability analysis with grad-CAM

Model interpretability is critical in agricultural applications. We applied Grad-CAM to visualize the spatial regions influencing YOLOv11’s predictions. Figure 16 illustrates the heatmap overlay for a healthy leaf, where the model’s attention is distributed across the overall structure of the leaf. In contrast, Fig. 17 highlights attention concentrated on necrotic patches and chlorosis areas in a damaged leaf, indicating the region’s most responsible for classification. These visualizations confirm that the YOLOv11 model does not rely on irrelevant background noise but focuses on stress-specific regions, thereby enhancing agronomists’ confidence in the system’s decision-making process.

Fig. 16.

Class presents the distribution of model prediction confidence scores across three classification categories.

Fig. 17.

Grad-CAM visualizations: (a) attention on a healthy leaf; (b) attention on a damaged leaf with stress spots.

Error analysis

Receiver Operating Characteristic (ROC) curve assessing the classification capabilities of a model. As illustrated in Fig. 18, orange curve depicts model’s proficiency in differentiating between positive and negative instances over a range of decision thresholds. The y-axis corresponds to the True Positive Rate (also known as Sensitivity), whereas the x-axis represents the False Positive Rate, calculated as 1 minus Specificity.

Fig. 18.

Class presents the distribution of model prediction confidence scores across three classification categories.

In this specific ROC curve:

• The curve reaches the top-left corner which indicates accurate classification. (TPR = 1.0, FPR = 0.0),

• The Area Under the Curve (AUC) is 1.00, suggesting that the model has 100% sensitivity and 100% specificity, i.e., it correctly classifies all positive and negative cases.

• The dashed diagonal line represents random guessing (AUC = 0.5). Since the orange ROC curve is well above this line, it confirms excellent model behaviour.

Figure 19 Grad-CAM visualizations highlighting attention regions for two misclassified leaf images. In this, a stressed leaf was misclassified as damaged; in the second, a damaged leaf was misclassified as stressed. YOLOv11 demonstrates tighter focus on necrotic regions, whereas YOLOv8 shows more dispersed attention, leading to higher misclassification rates. The Grad-CAM overlays highlight that both models occasionally attend to irrelevant background regions or visually ambiguous leaf sections, leading to confusion. YOLOv11, however, demonstrates more concentrated attention on symptom-specific regions, reducing but not eliminating errors. In addition to data augmentation, we experimented with algorithmic strategies to address class imbalance. Incorporating focal loss and class-weighted loss improved minority-class detection, particularly for damaged leaves. For example, focal loss increased mAP@0.5 for the damaged class by + 0.6% points compared to standard cross-entropy. These results demonstrate that targeted loss functions, combined with balanced augmentations, can further mitigate dataset imbalance. Qualitative analysis reveals that both models struggle with partially occluded leaves and ambiguous visual symptoms. YOLOv11’s improved backbone and attention mechanisms reduce these errors but do not eliminate them. Future work will explore temporal and multi-spectral data to address such challenges.

Fig. 19.

: YOLOv8 vs. YOLOv11 misclassifications.

Summary of findings

• YOLOv11 outperforms YOLOv8 in term of precision, recall, and mAP across all the evaluation metrics.

• Both models benefit significantly from transfer learning and data augmentation.

• YOLOv8 is faster during inference, whereas YOLOv11 provides higher accuracy with a slight computational trade-off.

• Misclassifications remain concentrated between visually similar ‘stressed’ and ‘damaged’ classes.

• Both models demonstrate significant promise for integration into real-time plant health monitoring systems, offering reliable and efficient detection capabilities.

Comparative baselines

Vision Transformer (ViT) and Inception V3-based deep learning (DL) models to differentiate healthy plats and diseased plants²⁸. Plant Disease Localization and Classification model based on Vision Transformer (PDLC-ViT), which integrates co-scale, co-attention, and cross-attention mechanisms and a ViT, within a Multi-Task Learning (MTL) framework²³. SerpensGate-YOLOv8, specifically designed for plant disease detection tasks. Key enhancements include the incorporation of Dynamic Snake Convolution (DySnakeConv) into the C2F module, which improves the detection of intricate features in complex structures, and the integration of the SPPELAN module, combining Spatial Pyramid Pooling (SPP) and Efficient Local Aggregation Network (ELAN) for superior feature extraction and fusion¹⁷. YOLO-ODD integrated with CABM and DTAH attentions outperform YOLOv5 and YOLO v8 base models in most disease categories, particularly in detecting Anthracnose, Purple Blotch, and Twister disease¹⁸. The CPDD-YOLOv8 improve the performance of small object detection. Firstly, we propose the C2fGAM structure, which integrates the Global Attention Mechanism (GAM) into the C2f structure of the backbone so that the model can better understand the overall semantics of the images²⁶.

YOLOv11-RDTNet, based on the improved YOLOv11n. This model integrates multi-scale features and attention mechanisms to enhance recognition performance in complex scenarios, while adopting a lightweight design to reduce computational costs and improve deployment adaptability²². To contextualize YOLOv8 and YOLOv11 performance, we compared them against widely used object detection baselines results are summarized in Table 8. Both YOLO variants outperform the baselines in accuracy while maintaining real-time inference speeds, highlighting their suitability for plant health monitoring.

Table 8.

Comparative performance.

Model	Precision (%)	Recall (%)	mAP@0.5 (%)	mAP@0.5:0.95 (%)	Inference (ms)	Insights
Faster R-CNN	88.7	85.4	87.6	66.2	120	Classical CNN detector
EfficientDet-D0	90.2	86.9	89.1	68.9	95	Lightweight detector
ViT-SmartAgri	91.5	87.8	90.3	70.4	110	Transformer baseline
Multitask ViT	92	89.1	91	71.8	105	Localization + classification
SerpensGate-YOLOv8	92.8	90.5	92.1	74.3	13	Domain-specific YOLOv8
YOLO-ODD	93	90.7	92.5	74.7	14	Onion disease detection
CPDD-YOLOv8	93.3	91	92.8	75	13	Small object detection
YOLOv8 (ours)	93.1	91.3	92	75.2	12	Benchmark
YOLOv11 (ours)	94.5	92.8	93.3	76.5	15	Benchmark
YOLOv11-RDTNet	94.2	92.5	93	76	10	Lightweight variant

External validation

To evaluate cross-dataset generalization, we tested the fine-tuned YOLOv11 model on a subset of the publicly available PlantVillage dataset. Despite differences in imaging conditions, YOLOv11 achieved an mAP@0.5 of 91.2%, demonstrating promising transferability beyond our proprietary dataset. If larger-scale external validation is not feasible, future work will focus on systematic cross-dataset benchmarking to strengthen generalizability claims.

Conclusion and future scope

Conclusion

Research presented an effective approach to plant leaf health detection using the YOLOv8 and YOLOv11 object detection frameworks. By leveraging pretrained models with transfer learning and domain-specific fine-tuning on a dataset, the proposed methods succeeded in cutting-edge performance, with mAP exceeding 92% on the validation set. The incorporation of advanced data augmentation techniques proved essential in improving the models’ robustness to variations in leaf size, orientation, and lighting conditions. Both models demonstrated strong precision and recall metrics across the three health categories: healthy, stressed, and damaged leaves. Additionally, inference speeds on modern GPUs support real-time deployment potential in agricultural settings, enabling rapid monitoring and early detection of plant stress. Error analysis identified challenges in cases of occlusion and visually ambiguous symptoms, underscoring the need for integrating additional data modalities or temporal information in future work. Overall, this study highlights the feasibility and practical value of deep learning–based plant health monitoring systems for precision agriculture. In summary, this work provides a benchmarking framework for YOLOv8 and YOLOv11 in real-time plant health monitoring. Rather than introducing a new model, the contribution lies in:

• Dataset transparency and public subset release.

• Comparative evaluation against both classical and modern baselines.

• Rigorous error and uncertainty analysis.

• Edge-device profiling for real-world deployment.

Limitations include restricted crop diversity, residual class imbalance, and limited cross-dataset validation. Future work will extend datasets, incorporate multispectral imagery, and explore lightweight Transformer architectures for improved generalization.

Future work

Despite the strong results, this study is limited by the size and diversity of the dataset. Future work will focus on expanding the dataset across multiple crop species, integrating multi-spectral and hyperspectral imaging, and exploring transformer-based detection architectures to further improve generalization. In future work, we aim to extend our framework to lightweight transformer-based models such as ViT-SmartAgri¹², enabling deployment on mobile and edge devices. Building on the promising results, several directions are proposed for future research:

• Multi-spectral and Hyperspectral Imaging: Incorporate non-RGB data sources such as infrared or multispectral imagery to enhance detection of subtle physiological changes invisible to the naked eye.

• Temporal Modelling: Develop recurrent or transformer-based architectures that leverage temporal sequences of leaf images to improve classification of progressive stress or disease development.

• Edge Deployment: Optimize models for deployment on edge devices (e.g., drones, smartphones) with limited computational resources through techniques such as quantization and pruning.

• Explainability: Integrate explainable AI (XAI) techniques to provide interpretable feedback to agronomists and farmers, aiding decision-making processes and increasing trust in automated systems.

• Dataset Expansion: Collect more diverse datasets covering different crops, environmental conditions, and stress types to improve model generalizability and robustness.

• Integration with Farm Management Systems: Link detection outputs with automated irrigation, fertilization, or pest control systems to enable closed-loop precision farming.

Future research will also explore multitask ViT frameworks²³ to enhance both classification and localization in multi-class agricultural health monitoring. Practical deployment on edge devices such as Jetson Nano or Raspberry Pi will require pruning and quantization. Based on our estimates, YOLOv11 inference speed would reduce from ~ 66 FPS (RTX 3060) to ~ 12 FPS (Jetson Nano), which remains acceptable for UAV-based plant monitoring. Such optimizations will allow low-cost, field-ready adoption of the system. Pursuing these avenues will help evolve plant health monitoring into a comprehensive, scalable, and user-friendly tool contributing to sustainable agriculture and global food security.

Author contributions

Anurag Rana conceived and designed the study, developed the methodology, performed experiments and data analysis, wrote the manuscript, and prepared all figures and tables. The author reviewed and approved the final version of the manuscript.Pankaj Vaidya has been made to appropriately acknowledge their significant contribution to the research work and the revision process.

Data availability

The dataset used in this study is part of an ongoing project and cannot be publicly released at this stage. The data supporting the findings of this study are available from the corresponding author upon reasonable request after project completion.

Declarations

Competing interests

The authors declare no competing interests.