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

2026-08-02

AI Summary

This work addresses the lack of efficient and reliable uncertainty estimation in agricultural semantic segmentation, where conventional ensembles incur high computational costs and single-model approximations often compromise quality. The authors propose ST-LoRA, the first approach to integrate Low-Rank Adaptation (LoRA) into ensemble learning for dense prediction. By training a shared, frozen backbone with lightweight, adapter-diverse members within a single training trajectory, ST-LoRA achieves both efficiency and diversity for uncertainty-aware segmentation. The study reveals that feed-forward layers are more effective than attention layers as LoRA adaptation targets, enabling the method to match or exceed full-rank ensemble performance with less than 10% of the parameters. On GrowliFlower-L and BUP20 datasets, ST-LoRA outperforms or matches existing efficient methods in segmentation accuracy, calibration, out-of-distribution detection, and cross-seed stability, while significantly reducing training time, inference latency, and storage overhead.

Abstract

Reliable decision-support in digital agriculture requires accurate predictions and well-calibrated uncertainty estimates, particularly for dense prediction tasks such as semantic segmentation. Ensemble methods provide strong uncertainty quantification, but their computational and memory demands limit practical use, while single-model approximations often trade off uncertainty quality for efficiency. We propose ST-LoRA, a parameter-efficient ensemble framework that builds 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 lightweight low-rank adapters, reducing trainable parameters to under 10% of the full model while preserving ensemble diversity. We evaluate across two agricultural datasets - GrowliFlower-L (cauliflower, open field) and BUP20 (sweet pepper, glasshouse) - using SegFormer and Mask2Former, covering in-distribution performance, calibration under distribution shift, and out-of-distribution detection. Ablations show feed-forward layers, not attention layers, are the critical LoRA target for dense prediction, contrary to the attention-only convention from language models. ST-LoRA matches or exceeds full-rank ensembles in segmentation accuracy and calibration across both datasets and architectures, while substantially reducing training time, inference latency, memory footprint, and storage requirements. Against efficient baselines - Snapshot Ensemble, MC Dropout, and Deep Deterministic Uncertainty - ST-LoRA consistently matches or outperforms them in image/pixel-level OoD detection, calibration stability under shift, and cross-seed variance, with far fewer parameters and lower compute. These results show LoRA-efficient ensemble adaptation is a highly effective, practical approach for uncertainty-aware agricultural vision systems.

Problem

• uncertainty quantification
• agricultural segmentation
• semantic segmentation
• efficient ensembles
• dense prediction

Innovation

• LoRA
• Snapshot Ensembling
• Uncertainty Quantification
• Parameter-Efficient Learning
• Agricultural Segmentation