An Efficient Attention-Gated Hybrid Transformer-CNN Framework for Plant Disease Segmentation In-the-Wild

Abstract

Abstract In real field scenarios in agriculture, automatic segmentation of plant diseases is an important technique for precision farming. However, it remains exceptionally challenging due to blurred lesions, complex morphological structures, irregular backgrounds, and severe class imbalance. While traditional convo lutional networks struggle to capture long-range semantic context and standard vision transformers fail to preserve sharp localized boundaries, this paper proposes an efficient, attention-gated hybrid framework optimized for field deployment. Our architecture leverages a hierarchical Mix Transformer (MiT-B2) encoder stream integrated with an Atrous Spatial Pyramid Pooling (ASPP) scale-space context bridge and a custom Cross-Scale Multimodal Attention Gate (CMAG) to isolate discriminative disease markers selectively. Evaluated on the highly challenging and unbalanced PlantSeg dataset, our framework achieves competitive mean Intersection over Union (mIoU) of 66.57% and an F1-score of 79.93%, while maintaining a highly compact parameter footprint of only 30.37 M. Experimental evaluations demonstrate that the proposed system establishes a new performance milestone, outperforming current competitive architectures and proving highly viable for resource-constrained edge devices. To further enhance out-of-distribution stability, we outline future directions to extend our top-performing candidate variants into a Level 1 meta-stacking ensemble optimized via few-shot learning and partial backbone fine-tuning.

Plant phenotyping relevance

植物病害領域の画像セグメンテーション手法を開発・評価し、病変の分割性能を定量検証しているため、植物の病害状態を推定するフェノタイピング手法が中心です。

Code and data availability

The paper uses the public PlantSeg dataset, but it is cited prior work ([29]), not an authors' deposit, and no allowed URLs exist to point to it. No author analysis code, trained model checkpoints, or data deposits are mentioned; the only supplementary file is a figure image (Aiincrops.png) with no stated phenotyping/`

No evidence-backed public reproduction asset is currently recorded.