Robust uncertainty-aware semi-supervised learning for agricultural vision

Develop robust semi-supervised learning frameworks for agricultural vision that effectively handle noisy labels, uncertain samples, and prediction uncertainty in plant disease detection.

Background

Plant disease detection systems often rely on limited labeled data and must exploit large collections of unlabeled images. Existing semi-supervised learning methods can suffer from error accumulation during pseudo-labeling and may treat uncertain predictions as reliable supervision. In agricultural vision, variations in disease appearance, environmental conditions, and visual similarity between classes make uncertainty handling particularly important.

The paper proposes U-SSDNet, an uncertainty-aware Mean-Teacher framework for sugarcane leaf disease detection that combines discrete wavelet preprocessing, KL-divergence-based pseudo-label refinement, and supervised contrastive learning. Although the proposed method addresses these issues for sugarcane leaf datasets, the broader challenge of achieving robust learning under noisy labels and uncertain predictions is explicitly identified as an unresolved research problem.