Gradient-Guided Boundary-Aware Selective Scanning with Multi-Scale Context Aggregation for Plant Lesion Segmentation

Guanqun Sun^1*Tianshuo Li²Yizhi Pan¹Zidan Zhu²Tianhua Yang²Feihe Shao²Jia Guo^3*Junyi Xin^2*

• ¹Japan Advanced Institute of Science and Technology, Nomi, Japan
• ²Hangzhou Medical College, Hangzhou, China
• ³Hosei University, Tokyo, Japan

Plant lesion segmentation aims to delineate disease regions at the pixel level to support early diagnosis, severity assessment, and targeted intervention in precision agriculture. The task remains challenging due to (i) large variation in lesion scale—from minute incipient spots to coalesced regions—and (ii) ambiguous, low-contrast boundaries that blend into healthy tissue. We present GARDEN, a Gradient-guided boundary-Aware Region-Driven Edge-refiNement network that unifies multi-scale context modeling with selective long-range boundary refinement. A Multi-Scale Context Aggregation (MSCA) module harvests contextual cues across diverse receptive fields to form scale-consistent lesion priors, improving sensitivity to tiny lesions while preserving large structures. A Boundary-aware Selective Scanning (BASS) module, conditioned on a Gradient-Guided Boundary Predictor (GGBP), produces an explicit boundary prior and uses it to steer a Mamba-based 2D selective scan—allocating long-range reasoning to boundary-uncertain pixels while relying on local evidence in confident interiors. Across two public plant disease datasets, GARDEN achieves state-of-the-art results on both overlap and boundary metrics, with pronounced gains on small lesions and boundary-ambiguous cases. Qualitative results further show sharper contours and reduced spurious responses to illumination and viewpoint changes. By coupling scale robustness with boundary precision in a single architecture, GARDEN delivers accurate and reliable plant lesion segmentation across challenging real-world conditions.