# A Comparative Evaluation of Deep Learning Architectures for Weed Classification, with an Exploratory Out-of-Distribution Analysis of Albanian Field Images

**Authors:** Rajesh Kumar (Department of Computer Science, University of New York Tirana, Tirana, Albania) — ORCID 0000-0001-8223-828X; Narasimha Rao Vajjhala (Department of Computer Science, American University in Bulgaria, Blagoevgrad, Bulgaria) — ORCID 0000-0002-8260-2392; Ervin Ramollari (Department of Computer Science, University of New York Tirana, Tirana, Albania) — ORCID 0009-0006-8614-7813; Earta Joca (Department of Computer Science, University of New York Tirana, Tirana, Albania)
**Type:** Journal article (open access)
**Source:** Information, 17(8), Article 751
**Published:** 2026-08-02
**DOI:** https://doi.org/10.3390/info17080751
**Canonical page:** https://www.narasimharao.net/research/deep-learning-weed-classification-deepweeds-albania/
**Indexing:** Scopus Q1 · Web of Science ESCI Q2 · Impact Factor 4.3 (JCR 2025)

## Summary

**Question.** Which modern deep learning architectures classify weed species most reliably, how stable are their rankings across random seeds, and do models trained on a public benchmark transfer to field images from Albania?

**Finding.** Across a multi-seed evaluation of six configurations on nine-class DeepWeeds data, fine-tuned DINOv2 and EfficientNet-B0 performed best overall — but the differences among several high-performing models were not consistently statistically significant across seeds. An exploratory test on unlabelled Albanian field images showed a clear geographic domain-shift problem.

**Why it matters.** Single-run leaderboard comparisons can overstate differences between models; this study shows why multi-seed evaluation, leakage-checked splits, and statistical testing matter for agricultural computer vision. For site-specific weed management in new regions such as the Western Balkans, locally collected and labelled data are needed before deployment.

## Key findings

1. The study compared six deep learning configurations — YOLO26n-cls, ViT-B/16, DINOv2 (ViT-S/14) with linear probing and full fine-tuning, ResNet-50, and EfficientNet-B0 — for nine-class weed classification on a stratified subset of the DeepWeeds benchmark.
2. Fine-tuned DINOv2 (DINOv2-FT) and EfficientNet-B0 showed the strongest overall performance.
3. Statistical analysis across repeated random seeds showed that differences among several high-performing models were not consistently significant.
4. Transfer-learning strategy and hyperparameter selection strongly affected model performance.
5. A reproducibility issue with ViT-B/16 was identified and reported, and the model was excluded from multi-seed statistical comparisons.
6. An exploratory out-of-distribution probe on unlabelled Albanian field images highlighted geographic domain shift and the need for locally collected, labelled datasets.

## Study at a glance

| Item | Detail |
|---|---|
| Research question | Which deep learning architecture is best for weed classification, and are the differences statistically reliable? |
| Models | YOLO26n-cls, Vision Transformer (ViT-B/16), DINOv2 (ViT-S/14) with linear probing and with full fine-tuning, ResNet-50, EfficientNet-B0 |
| Data | Stratified subset of the DeepWeeds benchmark, nine classes; unlabelled Albanian field images for an out-of-distribution probe |
| Methods | Leakage-checked train/validation/test split; repeated training across multiple random seeds; statistical comparison of models |
| Main result | DINOv2 (fine-tuned) and EfficientNet-B0 strongest; differences among top models not consistently significant across seeds |
| Implication | Local labelled datasets are needed before reliable regional deployment |

## Abstract

Automated weed classification supports site-specific weed management by enabling targeted interventions and reducing environmental and operational costs. This study presents a multi-seed evaluation of six deep learning configurations—YOLO26n-cls, Vision Transformer (ViT-B/16), DINOv2 (ViT-S/14) with linear probing and full fine-tuning, ResNet-50, and EfficientNet-B0—for nine-class weed classification using a stratified subset of the DeepWeeds benchmark. The models were evaluated using a leakage-checked train/validation/test split and repeated training across multiple random seeds to assess performance stability and statistical reliability. Among the evaluated approaches, DINOv2-FT and EfficientNet-B0 demonstrated the strongest overall performance, while statistical analysis showed that differences among several high-performing models were not consistently significant across seeds. The study also identifies the importance of appropriate transfer-learning strategies and hyperparameter selection, demonstrating that model performance can be strongly affected by optimisation choices. A reproducibility issue affecting ViT-B/16 was identified and transparently reported, leading to its exclusion from multi-seed statistical comparisons. An exploratory out-of-distribution probe using unlabelled Albanian field images further highlights the challenges of geographic domain shift and the need for locally collected, labelled datasets before reliable regional deployment. Overall, this work provides a systematic comparison of modern deep learning architectures for weed classification and emphasises reproducibility, statistical validation, and careful interpretation of model rankings.

## Key terms

- **Out-of-distribution (OOD) data:** Data that differ systematically from a model’s training data — here, field images from a different country, soil, and lighting than the benchmark.
- **Multi-seed evaluation:** Training and testing each model several times with different random initializations to check whether performance differences are stable rather than due to chance.
- **Linear probing vs. fine-tuning:** Linear probing trains only a new output layer on top of a frozen pretrained model; full fine-tuning updates all of the model’s weights.

## Limitations

- The Albanian field images were unlabelled, so the out-of-distribution analysis is exploratory.
- ViT-B/16 was excluded from multi-seed statistical comparisons because of a reproducibility issue.

## How to cite

Kumar, R., Vajjhala, N. R., Ramollari, E., & Joca, E. (2026). A Comparative Evaluation of Deep Learning Architectures for Weed Classification, with an Exploratory Out-of-Distribution Analysis of Albanian Field Images. Information, 17(8), 751. https://doi.org/10.3390/info17080751

```bibtex
@article{kumar2026deep,
  title = {A Comparative Evaluation of Deep Learning Architectures for Weed Classification, with an Exploratory Out-of-Distribution Analysis of Albanian Field Images},
  author = {Kumar, Rajesh and Vajjhala, Narasimha Rao and Ramollari, Ervin and Joca, Earta},
  journal = {Information},
  volume = {17},
  number = {8},
  pages = {751},
  year = {2026},
  publisher = {MDPI},
  doi = {10.3390/info17080751},
  url = {https://doi.org/10.3390/info17080751}
}
```