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

As deep learning models become standard in digital agriculture, providing reliable uncertainty estimates alongside predictions is critical for safety and decision-making. Traditional ensemble methods, which are the gold standard for uncertainty quantification, are often too computationally expensive for real-world deployment. This paper investigates whether Low-Rank Adaptation (LoRA) can be used to construct efficient, uncertainty-aware ensembles for dense prediction tasks like semantic segmentation, without the overhead of training multiple full-rank models.

The authors introduce ST-LoRA, a framework that combines LoRA with snapshot ensembling. Instead of training multiple independent models, ST-LoRA freezes a pretrained vision backbone and trains only lightweight low-rank adapters. By collecting multiple checkpoints from a single training trajectory, the authors create an ensemble of diverse adapters. They evaluate this approach on two agricultural datasets (cauliflower and sweet pepper) using SegFormer and Mask2Former architectures, testing performance under in-distribution conditions, distribution shifts, and out-of-distribution (OoD) scenarios.

ST-LoRA consistently matches or outperforms full-rank ensembles in segmentation accuracy and calibration while using less than 10% of the trainable parameters. A key discovery is that for dense prediction tasks, applying LoRA to feed-forward layers is more effective than the attention-only approach typically used in large language models. Furthermore, ST-LoRA demonstrates superior efficiency in training time, inference latency, and memory footprint compared to existing efficient baselines like MC Dropout and Deep Deterministic Uncertainty, making it a highly practical solution for agricultural vision systems.

This research provides a scalable path for deploying uncertainty-aware AI in agriculture. By reducing the computational barriers to high-quality uncertainty estimation, ST-LoRA allows for more reliable crop monitoring and disease detection systems that can flag when a model is uncertain, thereby reducing the risk of automated errors in critical agricultural workflows.