AgriSpot-Trans: Fine-Grained Micro-Lesion Segmentation in Complex Unstructured Agricultural Environments via Occlusion-Aware Hierarchical Transformer

International Journal of Engineering Intelligent Systems

Authors

  • Xianglin Qu Yunnan Open University, Kunming 650500,Yunnan, China
  • Haibo Peng Yunnan Open University, Kunming 650500,Yunnan, China
  • Xuanyi Zhu Yunnan Open University, Kunming 650500,Yunnan, China
  • Xiankang Shen Yunnan Open University, Kunming 650500,Yunnan, China
  • Xinran Chen Yunnan Open University, Kunming 650500,Yunnan, China
  • Jingyun Luo Yunnan Open University, Kunming 650500,Yunnan, China
  • Rong Zhou Baosight Software(Yunnan) Co., Ltd., Kunming 650500, Yunnan, China

Abstract

Accurate pixel-level segmentation of early-stage micro-lesions in unstructured agricultural environments is critical for precision agriculture, yet it remains fundamentally challenged by severe field occlusions, optical interferences, and fuzzy biological boundaries. To address these issues, we proposeAgriSpot-Trans, a novel end-to-end occlusion-aware hierarchicalTransformer. Unlike conventional hard-cropping pipelines that disrupt global canopy topology, our architecture introduces an Environment-AwareToken Routing (E-Decoupler) module. Driven by a Gumbel-Softmax mechanism, it softly suppresses background noise and specular reflections in the feature space. Furthermore, to overcome the morphological mismatch of standard
square-window attention, we design Morphological Radial Attention (MRA), which dynamically aligns its star-shaped receptive field with the radial diffusion gradients characteristic of biological tissue damage. Finally, a Level-Set Signed Distance Function (SDF) Decoder, optimized by a Boundary Aware Active Contour loss, is employed to achieve continuous, sub-pixel contour evolution. Extensive experiments on PlantVillage, RoCoLe, and our newly curated Complex-Field-Pest (CFP) benchmark demonstrate that AgriSpot-Trans achieves a state-of-the-art mIoU of 78.2% in complex field scenarios—a 9.5% absolute improvement over the strong SegFormer baseline (p < 0.01). Most notably, it boosts the recall for extreme micro-targets
(= 5 px) by an unprecedented 23.7%, delivering a highly robust and edge-efficient visual perception framework for real-time, variable-rate pesticide application.

Keywords: Precision agriculture, phytopathology, micro-lesion segmentation, vision transformer, level-set method, unstructured environments, active contour.

Cite As

X. Qu, H. Peng, X. Zhu, X. Shen, X. Chen, J. Luo, R. Zhou, "AgriSpot-Trans: Fine-Grained Micro-Lesion
Segmentation in Complex Unstructured Agricultural Environments via Occlusion-Aware Hierarchical
Transformer", Engineering Intelligent Systems, vol. 34 no. 3, pp. 289-303, 2026.


 

Downloads

Published

2026-05-01

Issue

Vol. 34 No. 3 (2026): International Journal of Engineering Intelligent Systems

Section

Article

License

The submission of a paper implies that, if accepted for publication, it will not be published elsewhere in the same form, in any language, without the prior consent of the publisher. Before publication, authors are requested to assign copyright to CRL Publishing Ltd. This allows CRL to sanction photocopying, and to authorize the reprinting of issues or volumes according to demand. Authors' traditional rights will not be jeopardized by assigning Copyright in this way, as they retain the right to reuse the material following publication, and to veto third-party publication.