ST-LoRA: Single-Trajectory LoRA Ensemble
for Uncertainty-Aware Agricultural Segmentation

Abstract

Reliable decision-support in digital agriculture requires not only accurate predictions but also well-calibrated uncertainty estimates, particularly for dense prediction tasks such as semantic segmentation. While ensemble methods provide strong uncertainty quantification, their computational and memory demands limit practical applicability, whereas single-model approximations often compromise uncertainty quality for efficiency. We propose ST-LoRA, a parameter-efficient ensemble framework that constructs diverse ensemble members from a single training trajectory by combining Low-Rank Adaptation (LoRA) with snapshot ensembling. Each member shares a frozen pretrained backbone and differs only in its lightweight low-rank adapters, reducing trainable parameters to less than 10% of the full model while preserving ensemble diversity. We conduct a comprehensive evaluation across two agricultural datasets — GrowliFlower-L (cauliflower, close-range/open field) and BUP20 (sweet pepper, close-range/glasshouse) — using SegFormer and Mask2Former architectures. Our evaluation covers in-distribution performance, calibration under distribution shift, and out-of-distribution detection. Extensive ablation reveals that feed-forward layers are the critical LoRA target for dense prediction — contrary to the attention-only convention established in language models. ST-LoRA matches or exceeds full-rank ensembles in segmentation accuracy and calibration across both datasets and architectures, while significantly reducing training time, inference latency, memory footprint, and storage requirements. Compared to efficient baselines including Snapshot Ensemble, MC Dropout, and Deep Deterministic Uncertainty, ST-LoRA consistently matches or outperforms them in image/pixel level out-of-distribution detection, calibration stability under distribution shift, and cross-seed variance, while requiring significantly fewer parameters and lower computational overhead. These results indicate that LoRA efficient ensemble adaptation is not merely a coarse approximation to full-rank ensembling, but rather a highly effective approach that offers a compelling efficiency-performance trade-off, making it a strong and practical competitor for uncertainty-aware agricultural vision systems.

Video

A short walkthrough of the method and findings.

Method

One training trajectory → many ensemble members.

• Single-trajectory snapshot ensemble. A cosine warm-restart schedule yields several snapshots along one run; the last four are ensembled (mean softmax) at inference.
• Parameter-efficient. Only low-rank LoRA adapters (plus the task head) are trained — the pretrained backbone stays frozen, so members share it and storage is tiny.
• Calibration-aware. Evaluated with float64 streaming ECE/ACE alongside mIoU and proper scores (Brier, NLL), avoiding float32 calibration artifacts.
• Feed-forward layers are the critical target. Our ablation shows that adapting the feed-forward layers — not attention alone — is what matters for dense prediction, contrary to the attention-only convention from language models.
• Architecture-agnostic. The same recipe applies to SegFormer and Mask2Former via collision-safe adapter placement on attention and feed-forward layers.

Experiments

Released as self-contained, reproducible experiments.

Datasets: BUP20 (sweet pepper) · GrowliFlower-L (cauliflower) · BUTom21 (tomato).

BibTeX

Copy@article{farag2026stlora,
  title   = {ST-LoRA: Single-Trajectory LoRA Ensemble for Uncertainty-Aware Agricultural Segmentation},
  author  = {Farag, Mohamed and Hoxha, Genc and Maleki, Yahya and McCool, Chris and Roscher, Ribana},
  journal = {arXiv preprint arXiv:2608.01530},
  year    = {2026}
}