PlantFormer: a precise plant disease segmentation network with interactive backbone and global-anisotropic context aggregation

Abstract

Precise plant disease segmentation in real-world agricultural environments presents challenges that general-purpose models often fail to address, primarily due to the anisotropic spread of lesions, blurred biological boundaries, and severe background dominance. To overcome these bottlenecks, this paper proposes PlantFormer, an end-to-end network that effectively integrates and adapts advanced architectural components to address these domain-specific issues. Specifically, PlantFormer employs an InteractSwin Backbone with a Cross-Level Fusion (CLF) module to preserve early pathological details. To model highly directional disease propagation, a GlobalAnisotropic Context Aggregation (GACA) neck utilizing strip pooling is introduced. Furthermore, a Semantic-Guided Fusion (SGF) decoder acts as a feature “boundary purifier” to suppress field noise, while a decoupled boundary-aware loss function explicitly shifts the optimization focus from healthy leaf regions to subtle necrotic transition zones. Comprehensive experiments demonstrate the effectiveness of our approach: PlantFormer achieves 41.78% mIoU on the complex PlantSeg dataset (unstructured field conditions) and 93.54% mIoU on the structured NLB dataset (vein-aligned lesions). It outperforms generalist models such as DeepLabV3+ and Segformer in key metrics like mIoU and mAcc. Despite these promising results, limitations remain, particularly regarding performance in scenarios with high-density, early-stage disease outbreaks, which will be the focus of future work.

Plant phenotyping relevance

植物病斑・壊死領域を直接セグメンテーションする画像解析手法の開発とベンチマークであり、植物病害状態の定量化が中心である。

Code and data availability

The supplied blocks describe PlantFormer evaluated on the public PlantSeg and NLB datasets, but these are cited third-party benchmark datasets, not paper-specific deposits. No data availability statement, code repository, trained model release, or author URL appears in the supplied text.

No evidence-backed public reproduction asset is currently recorded.