Robust plant disease segmentation in complex field environments: an in-depth analysis and validation with STAR-Net

Abstract

Introduction Plant disease segmentation in real-world agricultural environments poses significant technical challenges, including complex backgrounds, diverse lesion morphologies, and extreme class imbalance. Methods In this paper, we propose an integrated solution, STAR-Net, which combines a novel network architecture with a dynamic training strategy. The architecture features an innovative Heterogeneous Branch Attention Aggregation (HBAA) module to robustly represent multi-scale and multi-morphology features. The training strategy employs a Dynamic Phase-Weighted Loss (DPW-Loss) to navigate the complexities of imbalanced data. Results Our method achieves a state-of-the-art average mIoU of 93.36% on the NLB dataset. This result demonstrates its superior ability to precisely segment diseases with specific elongated morphologies. Furthermore, the model obtains a competitive average mIoU of 41.13% on the highly challenging PlantSeg dataset. This result validates its robustness in complex 'in-the-wild' scenarios. Discussion Our work presents a powerful, well validated, and synergistic solution for plant disease segmentation. It also paves the way for practical applications in precision agriculture.

Code and data availability

The supplied blocks describe the paper's methods and datasets (self-developed ADLD, plus public PlantSeg and NLB datasets from prior work), but contain no data availability statement, no authors' public code/model repository, and no deposited phenotype data or checkpoints. PlantSeg and NLB are cited third-party prior-d

No evidence-backed public reproduction asset is currently recorded.