SEAFEC: a spatial–edge adaptive convolution for multi-scale and boundary-aware plant disease and weed imagery

Abstract

Introduction Plant diseases and weeds are among the leading biological threats to global crop production. While deep learning has advanced automated analysis, existing approaches often fail under challenges like large multi-scale variations and blurred boundaries. Methods To address this, we propose SEAFEC (Spatial-Edge Adaptive Feature Enhancement Convolution), a novel convolutional module that jointly enhances scale adaptivity and boundary precision. SEAFEC employs a dual-branch design: the SCARF branch dynamically adjusts receptive fields, while the MEFE branch explicitly strengthens edge features. Results Across three representative tasks—plant disease classification, corn leaf disease detection, and sugarcane-weed segmentation—SEAFEC achieved consistent improvements (+1.8% accuracy, +2.5% mAP, +3.4% mIoU), with notable gains in boundary-sensitive cases. Discussion These results highlight SEAFEC as a general-purpose enhancement module, providing a unified solution for tackling scale-boundary challenges in agricultural imagery to support reliable disease diagnosis and precision weed management.

Plant phenotyping relevance

植物病害画像を対象に、マルチスケール・境界認識のための新規畳み込みモジュールSEAFECを開発し、病害分類・検出で技術性能を評価しているため、植物状態の画像ベース推定手法が中心である。

Code and data availability

The article uses public benchmarks (PlantVillage, corn disease, sugarcane-weed imagery) and YOLOv5/YOLOv8 frameworks, but provides no authors' public code, trained models, or paper-specific data deposit. The Data Availability Statement only offers supplementary material via the article and directs further inquiries to,

No evidence-backed public reproduction asset is currently recorded.