A Dataset-Centric Benchmark of Deep Learning Methods for Grape Leaf Disease Classification and Detection

• Controlled datasets can produce high classification accuracy while failing to represent field conditions involving complex backgrounds, illumination changes, occlusion, blur, viewpoint variation, and mixed disease symptoms.Models trained on controlled images may learn dataset-specific visual patterns rather than disease-specific features, reducing generalization to real vineyard imagery.Source: §1 · p. 3
• The study separates image-level classification, region-level classification, and object detection because complete images, annotated crops, and jointly localized targets represent different input units and annotation requirements.Region-level classification is used when annotations identify selected regions without sufficiently comprehensive coverage for conventional detection, whereas detection requires complete-image localization and categorization.Source: §1 · pp. 3–5
• Dataset properties—including acquisition conditions, provenance, class distribution, label compatibility, and annotation granularity—can substantially influence reported deep learning performance and cross-dataset reliability.The study emphasizes that overlapping disease labels do not guarantee equivalent recognition tasks because datasets may differ in target scale, symptom appearance, bounding-box definitions, and annotation density.Source: §1 · pp. 4–6
• Representative convolutional, transformer-based, hybrid, YOLO-family, and transformer-based models are evaluated with accuracy for classification and mAP@50 and mAP@50:95 for detection.The protocols are task-specific but internally consistent, with detection results interpreted separately because datasets define different semantic targets.Source: §1 · pp. 4–6
• The benchmark analyzes publicly available datasets by disease taxonomy, annotation type, acquisition conditions, image characteristics, class distribution, provenance, and suitability for distinct recognition tasks.This dataset-centric analysis supports practical interpretation of benchmark outcomes rather than architecture ranking alone.Source: §1 · pp. 4–5
• Cross-dataset experiments harmonize compatible disease categories for image-level classification and detection, while overlap analysis identifies exact image reuse among controlled or derivative datasets.The broader objective is to clarify how dataset shift, exact overlap, class imbalance, and annotation semantics affect robustness and benchmark interpretation.Source: §1 · pp. 5–6

2. Related Work

• The literature uses inconsistent terms for detection, identification, diagnosis, and recognition, requiring distinction among image-level classification, region-level classification, and complete-image object detection.Original terminology is retained for cited studies, while the experimental task is clarified according to whether localization is absent, supplied, or jointly predicted.Source: §2 · p. 6

2.1. Classical Image Processing and Machine Learning Approaches

• Classical methods used segmentation, handcrafted color or texture features, feature selection, and machine-learning classifiers, achieving reported accuracies of 88.89% and 93% while remaining sensitive to preprocessing and segmentation quality.Representative systems combined K-means or GrabCut segmentation with SVM, neural networks, AdaBoost, random forests, PCA, and Relief selection [18].Source: §2.1 · pp. 7–8

2.2. CNN and Transfer Learning Methods for Grape Leaf Disease Classification

• CNN and transfer-learning studies achieve high image-level classification performance, commonly using PlantVillage’s four-class grape subset, but mostly evaluate within one dataset and therefore provide limited evidence of generalization.Reported approaches include modified VGG16, MobileNet, AlexNet ensembles, VGG16-based CNNs, EfficientNetB0 features with Categorical Boosting, and ResNet50V2.Source: §2.2 · pp. 8–10

2.3. Lightweight, Attention-Based, Transformer-Based, and Hybrid Classification Models

• Lightweight, attention-based, transformer-based, and hybrid models target accuracy–efficiency tradeoffs and practical deployment, including MobileNet variants, GC-MobileNet, SwinGNet on Raspberry Pi, and optimized backbones.These studies increasingly address parameter reduction, interpretability, real-time operation, and edge deployment, but remain largely model-centric and single-dataset.Source: §2.3 · pp. 10–11

2.4. Grape Leaf Disease Detection and Localization

• Object-detection research extends recognition to spatial localization, but performance depends on annotation completeness and target semantics, which vary across vineyard datasets and may favor region-level classification.Studies include Faster R-CNN, attention-enhanced detectors, YOLO variants, mobile applications, LDD instance annotations, and HERMOS localized regions.Source: §2.4 · pp. 11–13

2.5. Dataset Bias, Disease-Label Heterogeneity, and Cross-Dataset Generalization

• Dataset bias, provenance, label heterogeneity, and incompatible annotation protocols make reported grape disease results difficult to compare and motivate dataset-centric cross-dataset evaluation.Controlled datasets can encode background and acquisition artifacts, while shared disease names do not ensure equivalent labels, targets, or detection tasks.Source: §2.5 · pp. 13–16

3.1. Image-Level Classification Datasets

• Image-level datasets span controlled, standardized, and realistic vineyard conditions, with major differences in taxonomy, resolution, scale, class balance, splits, and provenance.GVLiD and GL-Portugal preserve natural field variability, whereas NGLD uses resizing and background removal; GL-Portugal contains 5,267 field-acquired images, and its five classes comprise approximately 19.0%–21.4% each. NGLD is strongly imbalanced, with healthy leaves at 46.0%, downy mildew at 35.4%, and bacterial leaf spot at 3.7%.Source: §3.1 · pp. 19, 21, 23
• Disease-taxonomy differences support harmonized within-group and cross-dataset comparisons only for compatible class subsets.The common four-class configuration includes healthy leaves, black rot, Esca or black measles, and leaf blight or Isariopsis leaf spot, while NGLD [51], GLHCD [56], and GL-Portugal [58] add alternative categories such as downy mildew, powdery mildew, bacterial leaf spot, brown spot, and mite-related symptoms.Source: §3.1 · pp. 19–20, Table 2 · p. 20

3.2. Spatially Annotated Datasets and Benchmark Use

• Spatial annotations are assigned to region-level classification or detection based on target semantics and coverage, not annotation format alone.Region-level classification uses localized crops when annotations identify selected regions without exhaustive coverage, whereas object detection requires reliable matching of comprehensively annotated targets in complete images.Source: §3.2 · p. 25

3.2.1. Region-Level Classification Datasets

• HERMOS is used for region-level classification, yielding 13,852 regions from 490 source images across four classes with substantial class imbalance.The regions comprise 5,060 healthy, 1,882 dead-arm, 1,392 downy-mildew, and 5,518 powdery-mildew samples; incomplete annotation coverage makes the dataset unsuitable for conventional detection, and its accuracy is reported separately from image-level accuracy and detection mAP.Source: §3.2.1 · pp. 25–27

3.2.2. Object Detection Benchmark Datasets

• The object-detection benchmark uses three task-specific datasets whose targets differ between local symptoms and complete leaves: Mildew symptoms, filtered LDD, and FD-Confounders.Mildew contains 1,404 images and 24,775 annotations across two classes; filtered LDD contains 716 images and 9,231 annotations across four classes. Final benchmark versions reflect dataset-specific filtering, annotation conversion, duplicate checking, and split construction.Source: §3.2.2 · pp. 27–30, Table 3 · p. 28
• The datasets have limited category overlap: Mildew symptom and LDD share downy mildew and powdery mildew, whereas FD-Confounders uses Flavescence dorée, Esca, and confounding symptoms.The processed detection versions therefore do not define a common disease taxonomy, despite their shared grape-leaf recognition context.Source: §3.2.2 · p. 33, Table 4 · p. 33
• 28,038, 24,775, and 9,231 retained annotations comprise FD-Confounders, the Mildew symptom dataset, and the LDD leaf-only subset, respectively, with strong class imbalance across datasets.Downy mildew accounts for 84.3% of Mildew annotations; black rot accounts for 53.5% of LDD annotations, while grey mould represents approximately 0.7%. FD-Confounders contains 64.9% confounding leaves, 27.7% Flavescence dorée-symptomatic leaves, and 7.5% Esca-symptomatic leaves.Source: §3.2.2 · pp. 33–34, Figure 8 · p. 35
• Detection metrics should be interpreted within each dataset because annotation scale, class balance, scene complexity, annotation unit, and disease taxonomy differ substantially.Dominant classes can disproportionately influence aggregate mAP, while rare categories may be difficult to learn and evaluate reliably; absolute values should therefore not be treated as direct measures of task difficulty.Source: §3.2.2 · pp. 34–35
• Detection datasets provide complementary tasks: dense symptom-region detection, multi-class leaf-disease localization, and complete-leaf discrimination in complex whole-vine scenes.The Mildew symptom dataset emphasizes local symptoms, LDD supplies multiple disease categories with instance-level annotations, and FD-Confounders distinguishes target diseases from confounding symptoms. Healthy tissue is treated as background rather than a detectable object.Source: §3.2.2 · pp. 32, 33, 35

3.3. Dataset Characteristics and Benchmark Implications

• Benchmark tasks are assigned from annotation format and semantic completeness, with comparisons primarily within datasets and cross-dataset transfer restricted to compatible label spaces.Spatial annotations are not automatically considered exhaustive detection ground truth; transfer results are interpreted using acquisition conditions, provenance, image overlap, annotation semantics, and target granularity, especially for detection.Source: §3.3 · pp. 35–36
• Dataset partitions and aggregate results account for source-image grouping, duplicate content, class distribution, official protocols, and the limited supervision available for rare categories.Region-level classification is reported separately from complete-image classification because localization is supplied in advance.Source: §3.3 · p. 36

4. Benchmark Methodology

• The benchmark uses a unified framework to evaluate representative lightweight and moderately sized classification and detection models for resource-constrained vineyard monitoring.The design prioritizes practical deployment on mobile devices, embedded platforms, and edge-computing hardware rather than introducing a new architecture.Source: §4 · p. 36

4.1.3. Object Detection

• The three settings represent distinct practical objectives: complete-image recognition, discrimination after supplied localization, and localization plus classification in spatially annotated scenes.Object detection is suited to images containing multiple leaves, complex backgrounds, and spatially distributed symptoms, so results are interpreted separately.Source: §4 · p. 37, §4.1.3 · p. 38

4.1. Problem Formulation

• Three complementary evaluation settings are considered: image-level classification, region-level classification, and object detection.Together, they cover complete-image recognition, classification after manual localization, and joint localization and classification.Source: §4 · pp. 36–37, §4.1 · p. 37

4.1.1. Image-Level Classification

• Image-level classification predicts a disease or healthy class from each complete input image using a dataset-specific category set.Possible categories include healthy leaves, black rot, Esca or black measles, leaf blight, downy mildew, powdery mildew, bacterial leaf spot, brown spot, and mite-related symptoms.Source: §4.1.1 · p. 37

4.1.2. Region-Level Classification

• Region-level classification predicts labels for manually extracted annotated regions, evaluating discrimination after localization has already been provided.Partitions are created at the original-image level before crop extraction, preventing regions from one source image from crossing dataset splits.Source: §4.1.2 · pp. 37–38
• The region-level dataset contains healthy regions, dead arm, downy mildew, and powdery mildew categories.The model receives a manually localized region rather than the complete source image, so evaluation focuses on region discrimination.Source: §4.1.2 · p. 38

4.1.3. Object Detection

• Object detection predicts bounding boxes, class labels, and confidence scores for complete leaves, lesions, disease spots, or symptomatic regions.Only datasets with sufficiently consistent spatial annotations enter the quantitative benchmark, and available segmentation polygons are converted to enclosing bounding boxes.Source: §4.1.3 · p. 38

Model benchmark scope

• The benchmark covers residual, efficient, mobile-oriented, modern convolutional, transformer, hierarchical transformer, and hybrid CNN–Transformer architectures from timm.Table 5 summarizes their approximate parameter counts, GMACs, and intended comparative roles.Source: §4.2 · p. 39, Table 5 · p. 40

CNN architectures

• ResNet-18, ResNet-50, EfficientNet-B0, EfficientNet-B3, and MobileNetV3-Large compare residual capacity, compound scaling, and resource-efficient inference.The set contrasts lightweight and higher-capacity alternatives while reflecting deployment on smartphones, embedded boards, and other field devices.Source: §4.2 · pp. 40–41

Transformer and modern convolutional architectures

• ConvNeXt-Tiny, ViT-S/16, DeiT-S/16, and Swin-Tiny compare modern convolution with patch-based, data-efficient, and hierarchical attention mechanisms [65].These designs target local, distributed, multi-scale, and computationally efficient representation of disease patterns.Source: §4.2 · pp. 41–42

Hybrid architecture

• MobileViT-S combines convolutional local feature extraction with transformer-based long-range modeling for compact, edge-oriented recognition [68].Its inclusion bridges mobile CNNs and pure transformers where both compactness and global context may matter.Source: §4.2 · p. 42

Evaluation protocol

• Within each dataset and classification setting, every architecture uses identical partitions, preprocessing, augmentation, optimization, checkpoint selection, and evaluation metrics.Multiple datasets then allow acquisition conditions, class distributions, and provenance to be examined separately from architectural differences.Source: §4.2 · pp. 39, 42

4.3. Qualitative Classification Interpretation

• Grad-CAM overlays a coarse class-discriminative activation map on the input image to show regions contributing most strongly to the selected class score.The method weights convolutional feature maps according to target-class gradients.Source: §4.3 · p. 42
• ResNet-50 provides the representative visualization backbone because its final convolutional stage offers a direct, established target for Grad-CAM.This visualization backbone choice is independent of quantitative model ranking.Source: §4.3 · p. 42
• The analysis illustrates behavior on one correctly classified GL-Portugal test image and is not treated as symptom segmentation, localization ground truth, or an additional quantitative metric.It complements the quantitative image-level classification evaluation as a qualitative illustration only.Source: §4.3 · p. 42

Benchmark design

• Five detectors—YOLOv8n, YOLO11n, YOLO26n, YOLO26s, and RF-DETR-Nano—support comparisons of YOLO evolution, capacity scaling, and transformer-based detection.YOLOv8n, YOLO11n, and YOLO26n form a nano-scale generational comparison; YOLO26n versus YOLO26s isolates capacity; RF-DETR-Nano contrasts query-based set prediction with YOLO-style dense multiscale prediction.Source: §4.4 · pp. 43, 46

Evaluation scope

• Table 6 summarizes each detector’s approximate trainable parameters, GFLOPs for a 1,024 × 1,024 RGB input, and experimental role, while benchmark comparisons use mAP@50 and mAP@50:95.Models were profiled using fused inference configurations.Source: Table 6 · p. 44, §4.4 · p. 46

YOLO comparisons

• YOLOv8n, YOLO11n, and YOLO26n retain the backbone–neck–head single-stage structure but differ in feature extraction, attention, box regression, and inference procedures.These differences are examined at broadly comparable nano scale, with multiscale prediction suited to targets ranging from localized symptoms to complete leaves.Source: §4.4 · pp. 44–45

YOLO26 design

• YOLO26 introduces direct box regression, dual one-to-many and one-to-one heads, NMS-free inference, progressive loss weighting, and small-target-aware label assignment.These mechanisms target simpler deployment and datasets containing numerous small, spatially distributed symptom regions, but their practical benefit is evaluated empirically rather than assumed from architecture alone.Source: §4.4 · p. 45

Capacity scaling

• YOLO26n and YOLO26s share the same architecture but differ in network width and depth, with YOLO26s providing greater feature capacity at approximately four times the parameter count and computational complexity.This comparison tests whether additional representational capacity improves detection or whether the nano model is preferable for resource-constrained deployment.Source: §4.4 · p. 46

Transformer comparison

• RF-DETR-Nano uses a DINOv2 vision-transformer backbone, object queries, and matching-based set prediction, offering global-context reasoning distinct from YOLO dense prediction but at a substantially larger parameter scale than YOLO nano models.Its design may help with partial occlusion and complex vineyard scenes, while “Nano” is relative to the RF-DETR family rather than comparable in parameter count to YOLO nano models.Source: §4.4 · p. 46

Dataset Provenance and Image-Overlap Analysis

• Exact decoded-image hashing detects identical visual content despite differences in filenames, directory structures, or file encodings.Images are converted to a common pixel representation before comparison.Source: §4.5 · p. 47
• Overlapping images do not change within-dataset benchmark partitions but prevent substantial-overlap transfers from being treated as independent external validation.The analysis distinguishes genuinely independent source-to-target evaluation from transfer involving shared image content.Source: §4.5 · p. 47

Evaluation protocol

• Cross-dataset evaluation tests models trained on a source dataset directly on a target dataset without target-domain fine-tuning or adaptation, distinguishing transfer robustness from within-dataset performance.Within-dataset evaluation uses samples from the same processed dataset and acquisition and annotation protocol; the methodology also combines transfer with overlap analysis to assess robustness to independent images and annotation conventions.Source: §4.6 · pp. 47–48

Image-level classification

• Image-level transfer is restricted to datasets with compatible canonical disease categories, while derivative-dataset overlap is considered when interpreting external-validation evidence.The principal harmonized class set includes healthy leaves, black rot, Esca or black measles, and leaf blight or Isariopsis leaf spot, with particular attention to controlled-to-field transfer.Source: §4.6 · pp. 47–48

Object detection

• Detection transfer evaluates Mildew symptom and filtered LDD bidirectionally on downy mildew and powdery mildew using each dataset’s native annotation protocol.The experiment therefore combines visual-domain shift with annotation-protocol shift: direct symptom boxes in Mildew symptom versus boxes derived from local-symptom polygons in LDD.Source: §4.6 · p. 48
• Shared disease names are necessary but insufficient for detection transfer because the two datasets may define different visual targets through distinct annotation conventions.This interpretation follows from retaining native annotations while evaluating models across datasets.Source: §4.6 · p. 48

Region-level classification

• Region-level classification is evaluated only within its processed dataset because no second public dataset offers a sufficiently compatible region-level label space and annotation interpretation.Cross-dataset transfer is therefore not performed for this setting.Source: §4.6 · p. 48

5. Experimental Setup

• The benchmark separately evaluates complete-image classification, source-image-grouped region classification, and joint localization–classification detection using held-out test partitions.Classification uses complete images or manually localized annotated regions, while detection uses original complete images.Source: §5 · pp. 48–49, §5.1 · p. 49

5.1. Dataset Partitions

• Classification partitions use stratified 70%/15%/15% splits when needed, preserve official partitions when available, and prevent duplicate source files from crossing splits.HERMOS regions are grouped by source photograph, and detection partitions preserve dataset-specific or grouped category distributions.Source: §5.1 · pp. 49–50

5.2. Classification Pre-processing and Training Protocol

• Classification models use 224×224 ImageNet-normalized inputs, stochastic training augmentation, ImageNet initialization, and two-stage AdamW transfer learning with separate backbone and head learning rates.The backbone is first frozen, then jointly fine-tuned; validation loss determines checkpoint retention and early stopping.Source: §5.2 · p. 50

5.3. YOLO Training Protocol

• YOLOv8n, YOLO11n, YOLO26n, and YOLO26s are trained for 100 epochs at 1,024×1,024 resolution using pretrained Ultralytics checkpoints and validation-selected best models.YOLO26 evaluation uses the conventional NMS-based output path, with mAP extracted from bounding-box results.Source: §5.3 · pp. 51–52

5.4. RF-DETR-Nano Training Protocol

• RF-DETR-Nano uses directory adapters for original YOLO data, 1,024×1,024 inputs, effective batch size 16, EMA weights, and validation-based early stopping.Final predictions are evaluated through COCO-style bounding-box conversion and average precision over required IoU thresholds.Source: §5.4 · p. 52

5.5. Cross-Dataset Experimental Protocol

• Cross-dataset experiments train only on source data and evaluate directly on held-out target data, retaining harmonized labels and native detection annotations to expose domain and annotation shifts.Classification uses Swin-Tiny, while two-class detection transfers use YOLO26s without target-domain adaptation or model selection.Source: §5.5 · pp. 52–53

5.6. Qualitative Interpretability Setup

• Grad-CAM visualizes predicted-class attention for one correctly classified GL-Portugal test image using ResNet-50, independently of quantitative ranking or evaluation.The map is qualitative and is not interpreted as symptom segmentation.Source: §5.6 · p. 53

5.7. Evaluation Metrics

• Classification is evaluated with accuracy, whereas detection uses mAP@50 and mAP@50:95 computed independently per dataset to reflect differing taxonomies, scales, scenes, distributions, and annotations.mAP@50:95 averages precision across IoU thresholds from 0.5 to 0.95, providing stricter localization assessment.Source: §5.7 · pp. 53–54

6.1. Image-level Classification Results

• The image-level analysis used accuracy for within-dataset, class-level, overlap-supported, and compatible-label cross-dataset comparisons, with Swin-Tiny held constant across source–target experiments.Cross-dataset evaluation excluded NGLD, GLHCD, and GL-Portugal because their complete label spaces were incompatible with the common four-class configuration.Source: §6.1 · pp. 54–55, §6.1.4 · p. 59

6.1.1. Within-Dataset Accuracy

• Within-dataset accuracy is near-perfect for PlantVillage, PDR2018, the Grape Leaf Disease Dataset, its augmented version, and the New Plant Diseases grape subset, with all evaluated models reaching 1.0000 on the latter.GL-Portugal also performs highly at 0.9583–0.9836, whereas GVLiD peaks at 0.8966 and GLHCD at 0.8530, showing that acquisition and scene complexity strongly affect difficulty.Source: §6.1.1 · pp. 55–56, Table 7 · p. 55

6.1.2. Class-Level Accuracy on Challenging Datasets

• Class-level errors are concentrated in visually ambiguous categories: GVLiD struggles mainly with Esca/black measles and leaf blight, while GLHCD performs worst on brown spot and mites.Healthy leaves and black rot are classified correctly in all selected GVLiD test cases, whereas GLHCD’s lower accuracy reflects broader heterogeneity and symptom similarity.Source: §6.1.2 · pp. 57–58, Table 8 · p. 57

6.1.3. Dataset Overlap Analysis

• Exact image overlap links several controlled four-class datasets, including complete representation of PlantVillage within New Plant Diseases, so their near-perfect transfer cannot be treated as independent generalization evidence.No exact duplicates were detected between GL-Portugal and the other classification datasets.Source: §6.1.3 · p. 58, Table 9 · p. 58

6.1.4. Cross-Dataset Accuracy

• Cross-dataset accuracy exposes a major domain gap: models tested on GVLiD achieve only 0.1667–0.3218 from controlled or derivative sources, while GVLiD-trained models reach 0.2731–0.3250 on those targets.The contrast indicates that compatible disease labels do not ensure equivalent visual domains or robust transfer.Source: Table 10 · p. 59, §6.1.4 · p. 60
• The augmented dataset transfers only moderately, reaching 0.6098 on PlantVillage, 0.5810 on PDR2018, 0.6812 on the Grape Leaf Disease Dataset, and 0.6842 on New Plant Diseases, while controlled-source models reach 0.4224–0.6532 on it.Augmentation changes the visual distribution but does not fully reproduce field-acquired variability.Source: Table 10 · p. 59, §6.1.4 · p. 60

6.1.5. Summary of Classification Findings

• The benchmark’s overall interpretation is that accuracy depends jointly on acquisition environment, leaf visibility, occlusion, framing, background complexity, provenance, taxonomy, symptom separability, and class distribution.Consequently, reported accuracy should be accompanied by acquisition descriptions, overlap analysis, and cross-dataset evaluation when label spaces are compatible.Source: §6.1.1 · p. 56, §6.1.5 · p. 61

6.1.6. Qualitative Interpretation Using Grad-CAM

• Grad-CAM for a correctly classified GL-Portugal downy-mildew image assigned 92.9% probability and concentrated activation on the foreground leaf’s discolored symptomatic region rather than uniformly on the scene.The visualization is coarse class-discriminative evidence, not precise lesion segmentation or proof of consistent reliance on biologically meaningful evidence [76].Source: §6.1.6 · pp. 62–63, Figure 9 · p. 62

6.2. Region-Level Classification Results

• These accuracies measure discrimination on manually localized field-acquired regions, not complete-image recognition or automatic disease localization.The crops vary in dimensions, symptom scale, illumination, background content, and surrounding leaf tissue, so they should not be directly compared with Section 6.1.Source: §6.2 · pp. 63–64
• 0.9148 for DeiT-S/16 is the highest accuracy, versus 0.8837 for MobileViT-S, a 3.11-point range across all HERMOS region classifiers.Table 11 evaluates accuracy on the held-out HERMOS test partition, keeping regions from each source image in the same split to prevent leakage.Source: Table 11 · p. 63, §6.2 · p. 63
• Transformer models lead: DeiT-S/16 reaches 0.9148, Swin-Tiny 0.9138, and ViT-S/16 0.9054, with only 0.0010 separating the top two.These results indicate strong transformer performance for recognizing local symptom patterns within annotated regions.Source: §6.2 · pp. 63–64
• Convolutional models remain competitive, but scale is inconsistent: EfficientNetB0 reaches 0.9015, while EfficientNet-B3 underperforms it and ConvNeXt-Tiny reaches 0.8911.ResNet-50 improves over ResNet-18 from 0.8842 to 0.8975, showing a moderate benefit from depth without a universal capacity advantage.Source: §6.2 · p. 64

6.3. Object Detection Benchmark Results

• Detection uses mAP@50 for IoU 0.5 and mAP@50:95 averaged across IoU thresholds from 0.5 to 0.95, emphasizing increasingly precise localization.Across datasets, YOLO26s is the strongest YOLO-family model, while RF-DETR-Nano’s advantage is dataset-dependent rather than universal.Source: §6.3 · pp. 64, 66
• RF-DETR-Nano leads Mildew symptom and LDD, while YOLO26s leads FD-Confounders across the reported detection metrics.On Mildew, RF-DETR-Nano reaches mAP@50 0.7853 and mAP@50:95 0.5209; on LDD, 0.6114 and 0.4273; on FD-Confounders, YOLO26s reaches 0.5662 and 0.3109.Source: Table 12 · p. 65, §6.3 · pp. 65–66
• On Mildew symptom, YOLO variants are similar, YOLO26s is strongest among them, and RF-DETR-Nano further improves over YOLO26s by about 3.0 and 2.4 percentage points.Increasing YOLO26 from nano to small raises mAP@50 from 0.7363 to 0.7551 and mAP@50:95 from 0.4810 to 0.4972.Source: §6.3 · p. 65
• LDD shows progressive gains from YOLOv8n through YOLO26s, but RF-DETR-Nano remains best with mAP@50 0.6114 and mAP@50:95 0.4273.RF-DETR-Nano exceeds YOLO26s by only 0.0175 mAP@50 and 0.0113 mAP@50:95, producing comparatively close performance.Source: §6.3 · pp. 65–66
• FD-Confounders is hardest under precise localization, where YOLO26s reaches mAP@50:95 0.3109 and outperforms RF-DETR-Nano’s 0.2935.Its overlapping whole-vine scenes require dense localization and fine-grained discrimination among symptomatic and visually confounding leaves.Source: §6.3 · p. 66
• Best mAP@50:95 declines from 0.5209 on Mildew symptom to 0.4273 on LDD and 0.3109 on FD-Confounders, reflecting task differences rather than dataset quality.The datasets differ in annotation granularity, target scale, category count, scene complexity, and diagnostic similarity.Source: §6.3 · pp. 66–67

Experimental protocol

• The experiment used YOLO26s, retrained separately on the two shared categories and evaluated directly on the other dataset without target images, calibration, fine-tuning, or adaptation.YOLO26s was selected as the strongest YOLO-family detector in the within-dataset benchmark, while keeping architecture and training protocol fixed across directions.Source: §6.4 · p. 67, Table 13 · p. 68

Within-dataset references

• Within-domain references reached 0.3710/0.2079 mAP@50/mAP@50:95 on LDD and 0.7551/0.4972 on Mildew symptoms, confirming both models learned their source tasks.The source-domain gap is consistent with differences in annotation characteristics and detection difficulty, so transfer failure is not explained by unsuccessful two-class retraining.Source: §6.4 · pp. 68–69

Cross-dataset results

• 0.0010/0.0003 and 0.0016/0.0008 mAP@50/mAP@50:95 were obtained in the two transfer directions, retaining less than 0.3% of source-domain mAP@50.The LDD-trained model was evaluated on Mildew symptoms, while the Mildew-trained model was evaluated on LDD, with no target-domain adaptation.Source: §6.4 · p. 68, Table 13 · p. 68

Sources of transfer failure

• Identical disease labels did not define equivalent detection tasks because the datasets differ in box construction, target scale, annotation density, image conditions, backgrounds, and symptom appearance.LDD derives enclosing boxes from polygon annotations, whereas the Mildew symptom dataset directly boxes visible local symptoms, including many small distributed annotations in dense foliage.Source: §6.4 · pp. 67, 69

Class-distribution shift

• Class composition also shifts substantially: downy mildew forms 84.3% of Mildew annotations, while the LDD subset contains 2,292 downy-mildew and 1,933 powdery-mildew annotations.These differences can alter the learned decision boundary and detector confidence distribution under zero-adaptation transfer.Source: §6.4 · p. 69

Implications and limitations

• Reliable transfer requires compatibility in annotation semantics, target scale, bounding-box construction, acquisition protocol, and symptom appearance, not merely harmonized class names.Potential remedies include annotation harmonization, multi-dataset training, domain adaptation, or labelled target-domain fine-tuning; the present experiment serves as a stress test.Source: §6.4 · pp. 69–70

Benchmark scope

• The study evaluated image-level classification, region-level classification, and object detection separately because their input units, localization assumptions, annotation requirements, and practical interpretations differ.A Grad-CAM example assigned 92.9% probability to downy mildew and concentrated activation on the affected leaf region, but it was illustrative rather than precise segmentation or model ranking.Source: §7 · pp. 70–71

Image-level classification

• Controlled, augmented, or derivative datasets produced near-perfect or perfect classification accuracy, whereas GVLiD and GLHCD were harder because of field variability, scene complexity, taxonomy breadth, and visual heterogeneity.Exact image overlap among several controlled or derivative datasets further complicates interpreting apparent transfer, while transfer to GVLiD caused substantial accuracy degradation.Source: §7 · pp. 70–71

Region-level classification

• In region-level classification, DeiT-S/16 achieved 0.9148 and Swin-Tiny 0.9138 on HERMOS, while the narrow range indicated useful representations across convolutional and transformer models.These results apply after relevant regions were extracted and therefore should not be directly compared with complete-image classification or object detection; HERMOS was unsuitable for full-image detection.Source: §7 · pp. 71–72

Object detection

• RF-DETR-Nano led Mildew symptom and LDD detection with mAP@50:95 values of 0.5209 and 0.4273, while YOLO26s led FD-Confounders at 0.3109; no detector was universally superior.FD-Confounders was most demanding because it combines dense complete-leaf localization, overlapping vegetation, and fine-grained discrimination among diseases and confounding symptoms.Source: §7 · p. 71

Cross-dataset transfer

• YOLO26s cross-dataset detection collapsed between shared-label datasets: mAP@50:95 was 0.0003 from LDD to Mildew symptom and 0.0008 in reverse, versus source-domain values of 0.2079 and 0.4972.Thus, semantic label correspondence alone was insufficient to define a transferable detection task for this source–target pair.Source: §7 · pp. 71–72

Limitations and future work

• Reliable progress requires larger, diverse, expert-verified field datasets, standardized terminology and annotations, realistic protocols, and cross-dataset validation across cultivars, seasons, regions, and devices.The benchmark covers public datasets but does not establish clinical or agronomic validity across all growing conditions; future strategies include domain adaptation, annotation harmonization, and target-domain fine-tuning.Source: §7 · pp. 72–73