PaperlayerView exact revision on arXiv

Source-linked AI summary

When Ground-Truth Fidelity Matters: An Orchestrated UAS Framework for Wheat Streak Mosaic Virus Detection Using Vision Transformers and Machine Learning

Dewi Endah Kharismawati, Sandeep Dhakal, Courtney E. McCusker, Jennifer R. Wilson, Erik W. Ohlson, Sami Khanal

arXiv:2609.12169v1cs.CV

TL;DR

WSMV detection in sweet corn is complicated by variable symptoms, costly diagnostics, and label uncertainty. The paper develops a plant-level UAS multispectral workflow and evaluates it using labels of differing biological fidelity. Model performance changes substantially with label fidelity, and the study identifies limited ELISA-positive samples and a single site-season as constraints.

  • Problem

    Visual scouting can be unreliable and laboratory diagnostics are difficult to scale, while disease labels used for model training may not reflect plant-level infection status.

    Source: §1 · pp. 2–3

  • Method

    The study combines UAS imagery, plant-level processing, multispectral features, and Vision Transformer classification, comparing treatment-based labels with ELISA-confirmed labels.

    Source: §1 · p. 4

  • Results

    Model performance changed substantially as label fidelity increased, and adding NDVI and NDRE improved treatment-based classification.

    Source: §3.2 · p. 17, §4 · p. 22, §5 · p. 25

  • Takeaways & Limitations

    Scalable UAS-based WSMV phenotyping is technically feasible, but model interpretation depends critically on label fidelity.

    Source: §4 · p. 22, §4.5 · p. 25

    Evaluation was constrained by a single field site and growing season and a relatively small number of ELISA-confirmed positive samples.

    Source: §4.5 · pp. 24–25

Abstract

from arXiv · show

Wheat streak mosaic virus (WSMV) is a destructive pathogen of sweet corn and other cereal crops, causing yield losses and complicating early detection because symptoms are spatially variable and subtle. In sweet corn seed production, WSMV also has regulatory importance, as phytosanitary regulations from countries such as New Zealand and Chile require seed lots to be certified virus-free. Visual scouting is unreliable because symptoms can resemble abiotic stress, while enzyme-linked immunosorbent assay (ELISA) is accurate but expensive, labor-intensive, and difficult to scale. We present an automated pipeline for plant-level WSMV detection using unmanned aircraft systems (UAS) multispectral imagery. The framework integrates orthomosaic reconstruction, geospatial alignment, plant extraction, and classification using a Vision Transformer with seven-channel inputs (five spectral bands, NDVI, and NDRE). Using treatment-based labels, the model achieved 89% accuracy on over 6,500 test patches across multiple growth stages. However, ELISA-based ground truth revealed substantial label noise: only a small fraction of sampled plants in inoculated plots were infected. Treatment labels therefore did not reliably represent infection status, and the high accuracy was largely driven by label bias rather than disease detection. Performance decreased markedly against row-level symptom severity and plant-level ELISA labels. Under these higher-fidelity but smaller-sample conditions, both deep learning and classical machine learning showed limited generalization and weak separability between ELISA-confirmed mock-inoculated and infected plants. These results show that UAS-based disease detection is constrained by label fidelity and data availability, emphasizing biologically grounded labels and models aligned with real-world conditions.

1. Introduction

WSMV poses production and seed-trade risks, while current detection methods and plot-level remote sensing leave challenges for early, plant-level assessment. The study develops a plant-level UAS framework and evaluates how label fidelity affects model assessment.

1. Introduction

  • WSMV can cause yield and quality losses, disrupt mechanized sweet-corn processing, and create phytosanitary risks for seed production and international trade.Infection symptoms and their severity vary, and even low-level or spatially heterogeneous infections may threaten shipment acceptance or market access.Source: §1 · p. 2
  • Visual scouting can be unreliable, while accurate laboratory assays such as ELISA are costly, labor-intensive, and impractical at commercial seed-production scale.Symptoms can overlap with nutrient deficiencies, abiotic stress, or other disorders, especially early in infection.Source: §1 · p. 2
  • Plot- and canopy-level remote sensing can miss localized plant-level infection, while treatment, symptom, or plot labels may not match biological infection status.The paper identifies spatially heterogeneous WSMV infection and label discrepancies as challenges for model training and generalization.Source: §1 · p. 3
  • The study extends UAS orchestration from plot phenotyping to plant-level WSMV detection and compares treatment-based labels with ELISA-confirmed reference labels.Plant-level analysis brings additional challenges, including instance extraction, noise sensitivity, and the need for high-quality individual-plant labels.Source: §1 · p. 3
  • Vision Transformers can model spatial and spectral image patterns, but their use for plant-level field disease detection with sparse biological labels remains limited.The paper calls for disease-detection frameworks that combine scalable image-based modeling with explicit evaluation of label quality.Source: §1 · pp. 3–4
  • The study evaluates multispectral and vegetation-index features, label fidelity, and how label noise and limited high-fidelity labels affect model performance and generalization.Its end-to-end UAS workflow combines orthomosaic reconstruction, geospatial alignment, plant extraction, feature representation, and transformer classification.Source: §1 · p. 4

2. Materials and Methods

The field experiment used replicated WSMV-inoculated and mock-inoculated blocks containing multiple corn lines. Inoculation and ELISA sampling supported the study’s plant-level analyses.

2.1. Field Experimental Design

  • The 2025 experiment was conducted at the Schaffter Agricultural Research Station in Wooster, Ohio, in a field approximately 54.9 m wide and 274.3 m long.The soil was Wooster-Riddles silt loam.Source: §2.1 · p. 4
  • Six randomized blocks comprised three WSMV-inoculated and three mock-inoculated blocks, with 24 corn lines and three replicated rows per line in each block.Across the field, the design included 432 rows; resistant field-corn hybrids separated treatment blocks to reduce unintended virus spread.Source: §2.1 · p. 4, Figure 1 · p. 5
  • WSMV inoculations began around V4 and were applied four times; mock plants received buffer with carborundum rather than virus inoculum.The inoculum was prepared from infected corn tissue and applied with a mist blower.Source: §2.1 · p. 5
  • The structured field layout supported spatial sampling, and plant samples were collected for ELISA-based validation of infection status.Plots were machine-planted, and a pre-emergent herbicide mixture was used for weed control.Source: §2.1 · pp. 4–5

2.2. Data Acquisition

UAS and ground-based data were collected across five crop-growth stages. Ground control points and flagged plant samples supported spatial alignment between imagery and field observations.

2.2. Data Acquisition

  • Five UAS flights spanned V4 through reproductive growth stages, from June 20 to August 8, 2025.The dates were June 20, July 10, July 17, July 24, and August 8.Source: §2.2 · p. 5
  • RGB imagery was collected with a DJI Mavic 2 Pro, while multispectral imagery came from a DJI Matrice 200 carrying a MicaSense RedEdge-MX sensor.The multispectral sensor captured five bands: blue, green, red, red-edge, and near-infrared.Source: §2.2 · pp. 5–6
  • Imagery used 70% forward and side overlap, and five RTK-GPS-surveyed ground control targets supported georeferencing and alignment across dates.Four targets were near field corners and one near the center; the coordinates had centimeter-level positional accuracy.Source: §2.2 · p. 6
  • Plants were visually assessed and sampled for ELISA near UAS acquisition dates, with flags captured in imagery to identify sampled plants spatially.The figure depicts inoculation, RTK-GPS coordinate collection, and flagging of ELISA-sampled plants.Source: §2.2 · p. 6, Figure 2 · p. 6

2.3. ELISA Testing and Symptom Severity

Plant tissue was tested by ELISA and linked to UAS imagery, while symptom severity was scored by row. The section also specifies the threshold used to classify ELISA results.

2.3. ELISA Testing and Symptom Severity

  • Up to four plants per plot were sampled across four flight dates, and 576 samples were tested by ELISA and geolocated for linkage to UAS imagery.The samples comprised 286 plants from inoculated blocks and 290 from mock-inoculated blocks.Source: §2.3 · pp. 6–7
  • ELISA-positive status was assigned when a sample’s normalized value exceeded twice the healthy control average for its sweet-corn line and flight date.Values slightly below the 2× cutoff were interpreted cautiously because they could still indicate possible infection.Source: §2.3 · p. 7
  • Rows received symptom-severity scores from 0 to 4, with scores above zero classified as symptomatic and zero as asymptomatic.The scale ranged from no symptoms to severe mosaic, necrosis, or plant death.Source: §2.3 · p. 7
  • Figure 3 displays five raw reflectance bands alongside NDVI and NDRE, with a separate value scale for each channel.The caption describes the indices as highlighting plant vigor and canopy variability.Source: Figure 3 · p. 8

2.4. Data Generation

The datasets span seven-channel UAS imagery and treatment-based, row-symptom, and plant-level ELISA labels. Plant localization and data preparation differ by dataset, with a separate detector developed for deployment.

2.4. Data Generation

  • The three datasets pair seven-channel imagery with treatment labels, row-level symptom assessment, or plant-level ELISA infection status.The treatment-based dataset has the most plant-level samples, while the ELISA dataset uses laboratory-confirmed labels.Source: §2.4 · p. 8

2.4.1. Treatment-Based Dataset

  • Treatment-based samples tracked plants from V4 detections across later flights, then used 70:15:15 splits and rotated square crops for training and validation.Different counts reflected successfully detected and retained plants; test patches were not rotationally augmented.Source: §2.4.1 · p. 9

2.4.2. Row-Level Symptom Severity Dataset

  • Row-level positives came from inoculated rows with symptom scores above zero; block-based splits and manually selected mock rows addressed leakage and class imbalance.Rows were cropped and augmented, and ViT and ResNet18 models used the resulting dataset.Source: §2.4.2 · p. 10

2.4.3. Plant-Level ELISA Dataset

  • The ELISA dataset linked sampled plants to infection status, used block-level splits with mock samples selected to approximately match positives, and supported image and feature-based models.Classical models used spectral, vegetation-index, statistical, and GLCM texture features; image patches were used for deep learning.Source: §2.4.3 · pp. 10–11

2.4.4. Automated Plant Localization for Deployment

  • Dataset construction used prior seedling detections or sampled-plant GPS coordinates, neither of which can be assumed available for new imagery.A YOLOv26 detector trained on 664 augmented patches was developed to localize plants directly at the target flight date.Source: §2.4.4 · pp. 11–12

2.5. Vision Transformer Training for Plant-Level WSMV Classification

The study fine-tuned a Vision Transformer on seven-channel plant patches, using training-time augmentation and validation Macro F1 to select the final checkpoint.

2.5. Vision Transformer Training for Plant-Level WSMV Classification

  • The classifier used seven-channel plant patches resized to 224 × 224, with per-sample percentile clipping to reduce radiometric variability.The pretrained patch embedding was adapted from three input channels to seven by repeating the mean RGB kernel.Source: §2.5 · p. 12
  • Training used AdamW, label-smoothed cross-entropy, mixed precision, gradient clipping, and geometric augmentation, with up to 100 epochs and effective batch size 256.The classifier head learning rate was scaled 10× relative to the pretrained backbone.Source: §2.5 · p. 12
  • The best checkpoint was selected by validation Macro F1, and EMA weights were used for evaluation when they improved that score.Intermediate checkpoints were also saved every five epochs.Source: §2.5 · p. 12

2.6. Classical Machine Learning Classification

Classical classifiers used extracted plant-level ELISA features, with SVM preprocessing and feature selection tuned on validation Macro F1 and the held-out test reserved for final evaluation.

2.6. Classical Machine Learning Classification

  • Classical models used 65 spectral, vegetation-index, statistical, and GLCM texture features from the plant-level ELISA dataset, excluding image metadata.Source: §2.6 · p. 12
  • Linear and RBF SVMs used training-set standardization and ANOVA F-test feature selection, with feature counts and hyperparameters optimized on validation Macro F1.Source: §2.6 · pp. 12–13
  • XGBoost used extracted features directly, while PCA on standardized features visualized class separability; preprocessing and model selection used training and validation data only.The held-out test set was reserved for final evaluation.Source: §2.6 · p. 13

2.7. Orchestrated UAS phenotyping and Inference Pipeline

The adapted HPC-oriented workflow automates UAS image processing through classification and geospatial visualization. It preserves modular orchestration while adding seven-channel processing and transformer inference.

2.7. Orchestrated UAS phenotyping and Inference Pipeline

  • The workflow was designed to process multispectral imagery through aligned orthomosaics, raster stacks, patch extraction, inference, and geospatial outputs consistently across flight dates.Source: §2.7 · p. 13
  • The study adapted Waltz et al. (2025)'s YAML-configured, HPC-oriented pipeline with spectral-index calculation, model-ready patches, ViT inference, and spatial prediction visualization.Figure 4 distinguishes data generation, model development, and inference orchestration by swimlane and workflow-role colors.Source: §2.7 · p. 13, Figure 4 · p. 14
  • Orthomosaics were georeferenced and co-registered, then converted to seven-channel stacks for plot extraction and non-overlapping square patch tiling.The five spectral bands were aligned without resampling original pixel values; NDVI and NDRE completed the stack.Source: §2.7 · pp. 14–15
  • Inference produced patch-level predictions and confidence scores, with PNG overlays, shapefiles, and an interactive viewer supporting spatial inspection across dates.Performance metrics included precision, recall, F1 score, and accuracy.Source: §2.7 · p. 15

3. Results

Treatment-based patch classification and seven-channel inputs performed well, but evaluation against row-level symptoms and ELISA ground truth exposed label mismatch and limited generalization. Plant-level features showed substantial class overlap, and classical models performed weakly.

3.1. Patch-level classification using treatment-based labels

  • On 6,574 held-out patches, the treatment-labeled model achieved 89.25% accuracy and macro F1 0.8925; validation Macro F1 peaked at 0.8906 at epoch 53.Figure 5 depicts training and validation loss stabilization and the validation Macro F1 peak.Source: §3.1 · p. 16, Figure 5 · p. 16

3.2. Impact of seven-channel multispectral input

  • Accuracy rose from 0.841 to 0.892 and virus-labeled recall from 84.17% to 89.99% with seven channels; Figure 6 shows fewer off-diagonal errors for both classes.The seven-channel configuration added NDVI and NDRE to the five multispectral bands and reached its best validation performance at epoch 53, versus epoch 83 for five bands.Source: Figure 6 · p. 17, §3.2 · p. 17

3.3. Pipeline scalability and automated HPC execution

  • Across four flight dates, orthomosaic reconstruction dominated runtime; row-level metrics stayed within a narrow range despite variation in reconstruction time.The pipeline generated aligned orthomosaics, predictions, georeferenced files, and interactive visualizations through automated HPC execution.Source: §3.3 · pp. 17–18, Table 2 · p. 18, Figure 7 · p. 19

3.4. Effect of ground-truth refinement on label quality

  • ELISA confirmed 29 of 286 inoculated-block samples (10.1%), while treatment-trained predictions achieved 58–60% row accuracy and only 12–14% symptomatic precision.The sampled ELISA positives are confirmed infections, not a complete field census; Figure 8 illustrates the spatial comparison between predictions and symptom annotations.Source: §3.4 · p. 19

3.5. Row-level classification with Vision Transformer and ResNet18

  • ResNet18 correctly classified 57.8% of mock and 56.4% of symptomatic rows (symptomatic F1 0.52), versus 67% and 37% for the Vision Transformer, trading mock errors for symptomatic detection.Overall row-level performance remained below treatment-based patch-label performance, with heterogeneous symptom expression limiting class separability and generalization.Source: §3.5 · p. 20

3.6. Plant-level classification using ELISA ground truth

  • ELISA-grounded plant classification overfit, confused mock and virus plants, and overpredicted virus; few confirmed positives and weak image-feature correspondence constrained generalization.Validation precision, recall, and F1 remained unstable and did not improve consistently across epochs.Source: §3.6 · pp. 20–21

3.7.1. Feature Space Analysis Using PCA

  • PCA showed mock and virus samples substantially overlapping in both training and test projections, with no clear separation along the dominant components.The similar training and test distributions did not show a strong split shift in the dominant components; extracted spectral, vegetation-index, and texture features had limited discriminatory signal.Source: §3.7.1 · p. 21, Figure 9 · p. 21

3.7.2. Classification Performance Across Models

  • Linear SVM achieved test accuracy 0.558 and Macro F1 0.558, while RBF SVM achieved accuracy 0.558 and Macro F1 0.554; XGBoost accuracy was 0.404.The two SVMs performed similarly, while XGBoost also showed overfitting; substantial mock-virus misclassification remained across all three models.Source: §3.7.2 · pp. 21–22

4. Discussion

The discussion finds that UAS-based WSMV phenotyping is feasible, but performance depends strongly on whether labels represent individual-plant infection. It emphasizes spatially explicit processing, biological constraints on detectable symptoms, and the need for more scalable, higher-fidelity ground truth.

4.1. Technical Feasibility of Multispectral UAS-Based WSMV Detection

  • Seven-channel inputs incorporating NDVI and NDRE improved treatment-based classification, but treatment differences were not direct infection signatures because inoculation did not ensure individual-plant infection.Vegetation indices provided complementary information beyond raw multispectral bands.Source: §4.1 · p. 22

4.2. Scalable Orchestration for Multi-Date Disease Phenotyping

  • The automated multi-date pipeline produced spatially explicit patch predictions for comparison with row severity and ELISA, while orthomosaic reconstruction remained the main runtime bottleneck.Downstream processing, inference, and visualization contributed relatively little to total runtime.Source: §4.2 · p. 23

4.3. Label Fidelity and Biological Ground Truth

  • Treatment-based accuracy agreed less well with row-severity and ELISA-confirmed infection labels, suggesting the model learned treatment-associated patterns rather than consistently detecting plant-level WSMV.Row-level labels were spatially heterogeneous, and limited ELISA-positive samples and weak image-feature correspondence constrained generalization.Source: §4.3 · pp. 23–24
  • Row-severity labels were biologically more relevant but mixed symptomatic and nonsymptomatic plants, while classical image-derived features still overlapped substantially between mock and virus classes.The results point to difficulty capturing confirmed infection from imagery under sparse, heterogeneous field conditions.Source: §4.3 · pp. 23–24

4.4. Biological Constraints and Future Directions

  • Only three of 24 corn lines consistently showed field symptoms, likely weakening disease signals; image-based detection requires measurable canopy or spectral responses.The lines were IL110K, Oh28, and IL1171.Source: §4.4 · p. 24
  • Future work should increase ELISA-confirmed samples, use more consistently symptomatic genotypes, test across locations and seasons, and consider learning frameworks that combine label-certainty levels.Repeated disease assessments could also model symptom development over time and its relationship to image features.Source: §4.4 · p. 24

4.5. Limitations

  • The single-site, single-season study had few ELISA-positive samples, sparse and inconsistent symptoms, and no explicit model of temporal disease progression.These constraints limit assessment of generalizability and robust plant-level disease detection.Source: §4.5 · pp. 24–25

5. Conclusion

The conclusion presents a feasible UAS framework for sweet-corn WSMV phenotyping, while emphasizing that apparent performance depends on label fidelity. It reports weaker performance with biologically grounded labels and calls for scalable ground-truth strategies.

5. Conclusion

  • Treatment-based labels supported high classification accuracy, but row-severity and ELISA-confirmed labels showed that proxy labels did not consistently represent plant-level infection.Performance was substantially lower under row-level symptom-severity and plant-level ELISA-confirmed labeling conditions.Source: §5 · p. 25
  • The framework combined multispectral imagery, vegetation-index stacking, deep learning, and automated phenotyping to generate spatially explicit predictions across multiple flight dates.The pipeline processed data from raw UAS imagery through disease predictions.Source: §5 · p. 25
  • Adding NDVI and NDRE to raw multispectral bands improved treatment-based patch-classification performance.Source: §5 · p. 25
  • The conclusion identifies scalable strategies for biologically meaningful ground truth as necessary for future progress in field-scale disease phenotyping.High accuracy from indirect labels does not necessarily indicate accurate disease detection.Source: §5 · p. 25

CRediT authorship contribution statement

The contribution statement assigns software, validation, and formal analysis to D.E.K., while several study activities were shared across the author team.

CRediT authorship contribution statement

Supplementary Materials When Ground-Truth Fidelity Matters: An Orchestrated

The supplementary materials present dataset-preparation workflows for treatment-based, row-severity, and plant-level ELISA analyses, alongside model-training and test diagnostics.

Supplementary Materials

  • Figure S1 transfers seedling-stage plant detections to later flight dates, adjusts for canopy growth, and splits patches into training, validation, and test sets.The split ratio is 70:15:15.Source: Figure S1 · p. 31
  • Figure S2 uses segmented plant masks and row metadata to assign severity-derived labels, split row patches by block, and train ViT and ResNet18 classifiers.Evaluation uses held-out row-level test data.Source: Figure S2 · p. 31
  • Figures S3 and S4 describe plant-level ELISA workflows for deep learning on detected-plant image patches and classical models on extracted spectral, vegetation-index, and texture features.The classical workflow evaluates Linear SVM, RBF SVM, and XGBoost.Source: Figure S3 · p. 32, Figure S4 · p. 32
  • Row-level diagnostics show overfitting signs in both ViT and ResNet18; ViT had low virus-class recall, while ResNet18 still confused mock and virus rows.Figures S5 and S6 show training dynamics and normalized test confusion matrices.Source: Supplementary Materials When… · p. 33, Figure S5 · p. 33, Figure S6 · p. 34
  • Plant-level ResNet18 training was unstable, and its ELISA-labeled test predictions substantially confused classes, including mock plants predicted as virus.Figures S7 and S8 show training dynamics and the test confusion matrix.Source: Supplementary Materials When… · p. 34, Figure S7 · p. 34, Figure S8 · p. 35
  • XGBoost overfit the ELISA-derived features, and Linear SVM, RBF SVM, and XGBoost did not consistently separate mock and virus classes on test data.Figures S9 and S10 show boosting-round training dynamics and normalized test confusion matrices.Source: Supplementary Materials When… · p. 36, Figure S9 · p. 36, Figure S10 · p. 36

Related papers

Loading 2609.12169v1…