AgriVision:
Error-Driven Optimization
for Precision Agriculture

Vahe Gdlyan,  Yexishe Makuchyan,  Anahit Israelyan,  Armen-Aris Shahinyan

Computer Vision & Machine Learning Research  ·  2026

Dataset: Global Wheat Head Detection

Multi-institutional benchmark for agricultural object detection

3,422
Total Images
147K+
Bounding Box Annotations
1024×1024
Native Resolution
11
Research Institutions

Source Domains

arvalis_1 arvalis_2 arvalis_3 ethz_1 rres_1 usask_1 inrae_1 nau_1 uq_1 utokyo_1 utokyo_2

Dataset Characteristics

  • Multi-regional phenotypic variance across 4 continents
  • Annotations: [x, y, width, height] bounding boxes per wheat head
  • Density range: 1 – 89 heads per image
  • Photometric variance: full-sun, overcast, shadow conditions
  • Growth stages: mature golden-brown dominant, immature canopies sparse

Data Preprocessing Pipeline

From raw annotations to training-ready YOLO format

01  Raw CSV ingestion — master_metadata.csv (3.37 MB)

02  Parse bounding boxes from nested JSON arrays

03  Convert [x, y, w, h] → YOLO normalized [cx, cy, w, h]

04  Stratified train/test split — 80/20 ratio

     ├── Train: 2,738 images (train.csv — 2.69 MB)

     └── Test: 684 images (test.csv — 673 KB)

05  Export to YOLO directory structure

     ├── images/train/ + labels/train/

     └── images/val/ + labels/val/

06  Final dataset archive: 3.92 GB (yolo_dataset.zip)

Augmentation Strategy

  • Mosaic — 100% probability (4-image tile composition)
  • HSV-V jitter — 0.4 → 0.6 (photometric hardening)
  • FlipUD — 0.0 → 0.5 (nadir rotational invariance)
  • Mixup — 0.0 → 0.2 (canopy overlap simulation)

Coordinate Transform

cx = (x + w/2) / W
cy = (y + h/2) / H

Absolute pixel → normalized center coordinates [0, 1]

Sample Diversity & Domain Variance

Cross-institutional photometric and phenotypic variation

Dense mature canopy — golden phenotype

Sparse distribution — high contrast

Overexposure — photometric stress

Shadow region — luminance challenge

OOD Stress Case — Immature Green Canopy

56 combined errors · avg. confidence 0.44

The dataset's phenotypic bias toward mature golden-brown wheat creates a domain boundary for immature green canopies — establishing the model's irreducible error floor.

Baseline Model Performance

Initial deployment metrics — YOLOv8s at 640px resolution

0.950
mAP@50
0.974
Recall
0.877
Precision
0.569
mAP@50-95

YOLOv8s  ·  640px  ·  50 epochs  ·  Global Wheat Dataset

If the model achieves 0.950 mAP...

...why does the yield estimation pipeline overpredict by 4–8%?

Error Forensics

Systematic decomposition of baseline failure modes

1,092

Zero-IoU False Positives

The baseline network generated spurious detections concentrated in low-luminance shadow regions and dense canopy overlap zones, where recall maximization incentivized speculative predictions.

How do you penalize a network
for speculative detections?

The input distribution is fixed.
The loss gradient must be re-engineered.

Proposed Loss Formulation

Focal Boundary Loss with Adaptive Photometric Weighting

ℒtotal = λ1 · ℒfocal_boundary + λ2 · ℒphotometric_weighting

Focal Boundary Loss

Amplifies gradient magnitude for predictions at overlapping boundary regions, imposing an asymmetric penalty on bounding box proposals that intersect adjacent ground-truth instances.

Adaptive Photometric Weighting

Dynamically scales loss contribution as a function of local luminance variance, enforcing elevated false-positive penalties in photometrically ambiguous regions.

Ablation Study

Systematic hyperparameter search and false-positive decay analysis

  • λ1 held constant at boundary-optimized value
  • λ2 scaled logarithmically across 12 controlled trials
  • Optimal ratio determined at FP decay inflection point
  • Convergence achieved at epoch 35 via early stopping (30% compute savings)

Result: Median prediction confidence — 0.434 → 0.605 (+39.4%)

Should a model optimize for
benchmark metrics or operational reliability?

Operational Superiority over Metric Optimization

Comparative Evaluation

Baseline vs. Tuned — quantitative performance analysis

Metric Baseline (640px) Tuned (1024px) Δ
mAP@50 0.950 0.944 −0.006
Recall 0.974 0.906 −0.068
Median Confidence 0.434 0.605 +39.4%
Zero-IoU FP Count 1,092 0 ELIMINATED

Production Deployment

End-to-end inference pipeline — Hugging Face Spaces

01  High-resolution drone imagery ingestion

02  Tuned YOLOv8s inference at 1024px

03  Non-Maximum Suppression (conf + IoU)

04  Agronomic analytics computation

     ├── Yield estimation (t/ha)

     ├── Spatial uniformity (CV%)

     └── Revenue projection

05  Automated PDF executive report

Live: huggingface.co/spaces/gdlyanvahe/AgriVision-Engine

Known Limitations & Future Work

Identified Limitations

■

Motion Blur Sensitivity

High-velocity UAV traversal introduces edge deformation that degrades bounding box localization precision.

■

Domain Constraint

Network weights specialized for wheat canopy phenotypic topology. Transfer to alternative crop species requires supervised retraining.

Proposed Extensions

Temporal Consistency Module

Inter-frame smoothing to absorb motion artifacts across sequential captures.

Multi-Spectral Data Fusion

Integration of NDVI spectral indices for direct chlorophyll density correlation.

REST API Microservice

FastAPI-based inference decoupling for IoT and third-party application integration.

Thank you. We will now take questions.